Method for collecting training data for training ai model, and electronic device for performing same

Abstract

The present invention relates to a method for collecting training data for training an AI model, and an electronic device. The method may comprise: acquiring a plurality of pieces of candidate data that can be used for training the AI model, and extracting one or more characteristics for the respective pieces of candidate data; calculating similarity between one or more characteristics of at least one piece of representative data included in conventional training data used for training the AI model and one or more characteristics extracted for the respective pieces of candidate data; acquiring additional training data from among the plurality of pieces of candidate data on the basis of the calculated similarity and a similarity range of at least one piece of representative data; training an AI model by using the additional training data; and updating information about the training data on the basis of the additional training data.

Description

Method for collecting training data for training an AI model and an electronic device for performing the same

[0001] The present disclosure relates to a method and electronic device for collecting training data for learning an AI model. More specifically, the present disclosure relates to a method and electronic device for using and processing training data for learning an AI model by performing a similarity determination between features extracted from data and representative data.

[0002] With the advancement of technology and the utilization of AI models across various fields, diverse methods are being devised to effectively perform training tailored to each sector. Appropriate training is essential for the efficient use of AI models, making the selection of data to use for training a critical issue.

[0003] In fields where vast amounts of data are collected, such as wireless communication systems, utilizing all collected data for AI model training requires a large amount of storage space for data processing and consumes significant computing resources during training.

[0004] Accordingly, the present invention aims to provide a method to eliminate wasted data storage space and improve the learning speed of an AI model by selectively collecting data to efficiently train an AI model from a vast amount of data.

[0005] According to one embodiment of the present disclosure, a method for collecting training data for training an AI model is provided. The method may acquire a plurality of candidate data that can be used for training an AI model. The method may extract one or more features for each of the plurality of candidate data. The method may calculate a similarity between one or more features of at least one representative data included in the existing training data used for training the AI ​​model and one or more features extracted for each of the plurality of candidate data. Based on the calculated similarity and the similarity range of at least one representative data, the method may acquire additional training data among the plurality of candidate data. The method may train an AI model using the additional training data. Based on the additional training data, the method may update information regarding the training data.

[0006] According to one embodiment of the present disclosure, an electronic device for collecting training data for learning an AI model may include a memory for storing a plurality of instructions and at least one processor for executing a plurality of instructions stored in the memory. By executing a plurality of instructions individually or collectively by at least one processor, the electronic device may acquire a plurality of candidate data that can be used for learning an AI model. By executing a plurality of instructions individually or collectively by at least one processor, the electronic device may extract one or more features for each of the plurality of candidate data. By executing a plurality of instructions individually or collectively by at least one processor, the electronic device may calculate a similarity between one or more features of at least one representative data included in existing training data used for learning an AI model and one or more features extracted for each of the plurality of candidate data. By executing a plurality of instructions individually or collectively by at least one processor, the electronic device may acquire additional training data among the plurality of candidate data based on the calculated similarity and the similarity range of at least one representative data. An electronic device can train an AI model using additional training data by executing a plurality of instructions individually or collectively by at least one processor. An electronic device can update information regarding training data based on additional training data by executing a plurality of instructions individually or collectively by at least one processor.

[0007] FIG. 1 is a diagram illustrating the operation of collecting training data for learning an AI model according to one embodiment, expressed in multiple modules.

[0008] FIG. 2 is a diagram showing a specific operation of a method for collecting training data for learning an AI model according to one embodiment.

[0009] FIG. 3 is a diagram illustrating the operation of collecting training data for learning an AI model according to one embodiment.

[0010] FIG. 4 is a flowchart illustrating the operation of collecting training data for learning an AI model according to one embodiment.

[0011] FIG. 5 is a diagram illustrating an operation for processing data based on similarity according to one embodiment.

[0012] FIG. 6 is a diagram illustrating the operation of adding representative data as an example of updating information regarding training data according to one embodiment.

[0013] FIG. 7 is a diagram illustrating an operation to update a similar range as an example of an information update regarding training data according to one embodiment.

[0014] FIG. 8 is a diagram illustrating the operation of deleting data according to an information update regarding training data according to one embodiment.

[0015] FIG. 9 is a diagram relating to example data for specifically explaining the operation of collecting training data for learning an AI model according to one embodiment.

[0016] FIG. 10 is a schematic block diagram of an electronic device according to one embodiment.

[0017] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0018] The present disclosure is capable of various modifications and may have various embodiments, and specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the embodiments of the present disclosure, and it should be understood that the present disclosure includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the various embodiments.

[0019] In describing the embodiments, detailed descriptions of related prior art are omitted if it is determined that such descriptions would unnecessarily obscure the essence of the present disclosure. Furthermore, numbers used in the description of the specification (e.g., first, second, etc.) are merely identifiers to distinguish one component from another.

[0020] The terms used in the embodiments of this specification have been selected to be as widely used as possible, taking into account the functions of this disclosure; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been arbitrarily selected by the applicant, and in such cases, their meanings will be described in detail in the description section of the relevant embodiments. Therefore, terms used in this disclosure should be defined not merely by their names, but based on their meanings and the content throughout this disclosure.

[0021] The scope of the present disclosure may be defined by the claims set forth below rather than by the detailed description above. Various features mentioned in one claim category of the present disclosure (e.g., in method claims) may also be claimed in other claim categories (e.g., in system claims). Furthermore, an embodiment of the present disclosure may include not only combinations of features specified in the appended claims but also various combinations of individual features within the claims. The scope of the present disclosure should be interpreted as including all modifications or variations derived from the meaning and scope of the claims and their equivalents.

[0022] In addition, components expressed as '~part (unit),' 'module,' etc. in this disclosure may consist of two or more components combined into a single component, or a single component may be divided into two or more components according to more detailed functions. These functions may be implemented in hardware or software, or through a combination of hardware and software. Furthermore, each component described below may additionally perform some or all of the functions performed by other components in addition to the primary function it is responsible for, and it is obvious that some of the primary functions performed by each component may be exclusively performed by other components.

[0023] Singular expressions may include plural expressions unless the context clearly indicates otherwise. Terms used herein, including technical or scientific terms, may have the same meaning as generally understood by those skilled in the art as described in this specification.

[0024] Throughout this disclosure, unless specifically stated otherwise, "or" is inclusive and not exclusive. Accordingly, "A or B" may mean "A, B, or both" unless clearly indicated otherwise in the context. In this disclosure, the phrases "at least one of" or "one or more of" may mean that different combinations of one or more of the listed items may be used, or that only any one of the listed items is required. For example, "at least one of A, B, and C" may include any of the following combinations: A, B, C, A and B, A and C, B and C, or A and B and C.

[0025] It will be understood that each block of the process flow diagrams and combinations of the flow diagrams can be executed by computer program instructions. Since these computer program instructions can be loaded into the processor of a general-purpose computer, a specialized computer, or other programmable data processing equipment, the instructions executed through the processor of the computer or other programmable data processing equipment create means to perform the functions described in the flow diagram block(s). Since these computer program instructions can also be stored in computer-available or computer-readable memory that can be directed toward the computer or other programmable data processing equipment to implement the function in a specific way, the instructions stored in such computer-available or computer-readable memory can also produce a manufactured item containing the means of instruction to perform the function described in the flow diagram block(s). Since computer program instructions can be loaded onto a computer or other programmable data processing equipment, instructions that perform a series of operation steps on the computer or other programmable data processing equipment to create a process executed by the computer can also provide steps for executing the functions described in the flowchart block(s).

[0026] Additionally, each block may represent a module, segment, or part of code containing one or more executable instructions for executing a specified logical function(s). It should also be noted that in some alternative execution examples, the functions mentioned in the blocks may occur out of order. For instance, two blocks described in succession may actually be executed substantially simultaneously, or the blocks may be executed in reverse order according to their corresponding functions.

[0027] When a part of a specification is described as "including" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "...part" or "module" as used in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or as a combination of hardware and software.

[0028] Embodiments of the present disclosure are described below in detail with reference to the attached drawings so that those skilled in the art can easily implement them. However, the present disclosure may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present disclosure in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0029] Functions related to artificial intelligence according to the present disclosure are operated through a processor and memory. The processor may be composed of one or more processors. In this case, the one or more processors may be general-purpose processors such as CPUs, APs, and DSPs (Digital Signal Processors), graphics-dedicated processors such as GPUs and VPUs (Vision Processing Units), or artificial intelligence-dedicated processors such as NPUs. The one or more processors control the processing of input data according to predefined operation rules or artificial intelligence models stored in memory. Alternatively, if the one or more processors are artificial intelligence-dedicated processors, the artificial intelligence-dedicated processors may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0030] The predefined rules of operation or artificial intelligence models are characterized by being created through learning. Here, being created through learning means that a predefined rules of operation or artificial intelligence models configured to perform desired characteristics (or objectives) are created by a basic artificial intelligence model being trained using multiple learning data by a learning algorithm. Such learning may be performed on the device itself where the artificial intelligence according to the present disclosure is executed, or it may be performed through a separate server and / or system. Examples of learning algorithms include supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning, but are not limited to the examples described above.

[0031] An artificial intelligence model may be composed of multiple neural network layers. Each of the multiple neural network layers has multiple weight values ​​and performs neural network operations through operations between the results of previous layers and the multiple weights. The multiple weights possessed by the multiple neural network layers can be optimized based on the learning results of the artificial intelligence model. For example, the multiple weights may be updated so that the loss value or cost value obtained from the artificial intelligence model during the learning process is reduced or minimized. Artificial neural networks may include deep neural networks (DNNs), such as Convolutional Neural Networks (CNNs), Deep Neural Networks (DNNs), Recurrent Neural Networks (RNNs), Restricted Boltzmann Machines (RBMs), Deep Belief Networks (DBNs), Bidirectional Recurrent Deep Neural Networks (BRDNNs), or Deep Q-Networks, but are not limited to the examples mentioned above.

[0032] The terms used in this disclosure will be briefly explained, and an embodiment of the present invention will be described in detail.

[0033] The terms described below are defined considering their functions in the present invention, and these may vary depending on the intentions or practices of the user or operator. Therefore, their definitions should be based on the content throughout this specification.

[0034] In the present disclosure, "training data" may refer to data for training a neural network model or an AI model to perform operations according to a specific purpose. In one embodiment of the present disclosure, "training data" may include, but is not limited to, network data, and may refer to any type of data for training or learning an AI model.

[0035] In the present disclosure, 'candidate data' may refer to data collected to train a neural network model or an AI model, and may mean a data set for which it is determined whether to use it as training data based on similarity between the data.

[0036] In the present disclosure, 'feature' refers to an individual attribute or characteristic of data, which may mean a unit processed as input by a learning algorithm. In one embodiment, a feature may refer to a value or variable that describes each data point within a data set. For example, in network data, it may include, but is not limited to, variables related to channels, traffic, and network environments, such as CQI, RI, number of terminals, transmission size, etc.

[0037] In the present disclosure, 'representative data' refers to data used to filter training data for learning a neural network model or an AI model, and may mean data that represents existing ranges of training data.

[0038] In the present disclosure, 'similarity' may refer to a criterion for quantitatively measuring the degree of correlation or agreement between data. In one embodiment of the present disclosure, 'similarity' may refer to a criterion for measuring correlation between representative data and candidate data, and may refer to a value that quantifies the difference for one or more characteristics. In one embodiment, 'similarity' may refer to a set of similarity values ​​measured for each characteristic.

[0039] In the present disclosure, the 'similarity range' refers to a range that a neural network model or an AI model can predict based on existing training data, and may mean a range of input data that a trained model can infer with a certain level of accuracy or higher. In one embodiment of the present disclosure, candidate data within the 'similarity range' of representative data may refer to data that, when input to an AI model, can appropriately infer an output value based on existing training data.

[0040] In the present disclosure, 'filtering' may refer to a processor that selects data from a data set according to specific criteria. In one embodiment of the present disclosure, training data may be filtered according to a similarity calculated based on the similarity range of representative data among candidate data.

[0041] For the sake of convenience of explanation, the following description focuses on network data, which is one embodiment of data for training an AI model. However, all embodiments described in this disclosure may be applied not only to network data but also to all types of data that can be trained or processed by an AI model.

[0042] FIG. 1 is a diagram illustrating the operation of collecting training data for learning an AI model according to one embodiment, expressed in multiple modules.

[0043] In one embodiment of the present disclosure, an electronic device (100) may obtain additional training data (120) for training an AI model from candidate data (110) and may update information regarding the training data based on the candidate data (110). In the present disclosure, training data refers to data for training an AI model to perform an action according to a specific purpose. Candidate data refers to data collected for training an AI model, which can be used as training data or filtered.

[0044] The electronic device (100) can obtain additional training data (120) by extracting one or more features from candidate data (110) and filtering them through the calculation of similarity with representative data. In the present disclosure, representative data refers to data that represents existing training data ranges. The electronic device (100) can update information regarding training data using candidate data (110).

[0045] Information regarding training data may include information regarding representative data and a similarity range for each characteristic of the representative data. In this disclosure, the similarity range refers to a range that an AI model can predict based on existing training data, and may refer to a range of data that a trained AI model can infer with an accuracy of a certain level or higher. In one embodiment, the similarity range may be set for one or more characteristics for each representative data. Specific examples are described below.

[0046] The operation of the electronic device (100) collecting training data can be described as a software unit responsible for a specific function or role. The modules (130 to 170) illustrated in FIG. 1 may be classified according to function as operations performed by executing instructions stored in memory in a processor. Therefore, the operations described below as being performed by the modules (130 to 170) illustrated in FIG. 1 can actually be seen as being performed by the processor.

[0047] In one embodiment of the present disclosure, the electronic device (100) may be an electronic device capable of processing and outputting a video or image. The electronic device (100) may be implemented in various forms including a display. For example, the electronic device (100) may be implemented as various electronic devices such as a TV, mobile phone, tablet PC, digital camera, camcorder, laptop computer, desktop, e-book reader, digital broadcasting terminal, PDA (Personal Digital Assistants), PMP (Portable Multimedia Player), navigation, MP3 player, wearable device, etc.

[0048] In one embodiment, an electronic device (100) may collect candidate data (110) containing network data for training an AI model in a wireless communication system. The electronic device (100) may select additional training data (120) from the candidate data (110) for efficient training of the AI ​​model and use it for training. By selecting only the additional training data (120) that cannot be predicted by the previously collected data from a vast amount of candidate data (110), the electronic device (100) can eliminate wasted data storage space and improve the training speed of the AI ​​model.

[0049] The electronic device (100) can identify additional training data (120) from candidate data (110) and update information regarding the training data through the operation of a plurality of modules (130 to 170). The electronic device (100) can perform training of an AI model using the additional training data (120) identified from the candidate data (110). The electronic device (100) can perform efficient training of an AI model using additional training data (120), which are unpredictable data based on existing training data.

[0050] In one embodiment of the present disclosure, an electronic device (100) may acquire a plurality of candidate data (110) that can be used to train an AI model. The candidate data (110) refers to data collected to train an AI model, which can be used as training data. In one embodiment, the plurality of candidate data (110) may include additional training data (120) and data that is filtered out and not used as additional training data (120).

[0051] In one embodiment, a plurality of candidate data (110) may include data generated in various network environments for training an AI model in a wireless communication system. The plurality of data (110) may include data generated by various regions, times, and operators.

[0052] The electronic device (100) can extract one or more characteristics for a plurality of candidate data (110). In one embodiment, the electronic device (100) can extract one or more characteristics from the candidate data (110) through a characteristic extraction module (130). The characteristic extraction module (130) can extract characteristics representing the attributes of the data according to the type of candidate data (110).

[0053] In one embodiment, one or more characteristics may include scalar characteristics or vector characteristics. For example, the characteristic extraction module (130) may extract characteristics regarding network data obtained from a wireless communication system.

[0054] The characteristics of network data are scalar characteristics and may include, but are not limited to, the number of terminals, transmitted bytes, received bytes, etc. The characteristics of network data are vector characteristics and may include, but are not limited to, CQI (Channel Quality Indicator), which represents quality information of the wireless channel, or RI (Rank Indicator), which represents the number of spatial streams in the antenna system.

[0055] The electronic device (100) can extract one or more features from at least one representative data. The representative data may refer to data that represents the range of existing training data used for training an AI model. In one embodiment, the electronic device (100) can extract one or more features from the representative data through a feature extraction module (130).

[0056] In one embodiment, the electronic device (100) may extract one or more characteristics of representative data and store them in a database prior to acquiring candidate data (110). The electronic device (100) may use one or more characteristics of representative data stored in a database (not shown).

[0057] The electronic device (100) can calculate the similarity between representative data and candidate data (110) using a similarity calculation module (140). The electronic device (100) can calculate the similarity between one or more characteristics of representative data and one or more characteristics of candidate data (110) using a similarity calculation module (140). In one embodiment, similarity may refer to a value that quantifies the difference between characteristics. Similarity may refer to a matrix containing similarity values ​​between one or more characteristics.

[0058] The electronic device (100) can calculate the similarity between representative data and candidate data for each characteristic extracted through the characteristic extraction module (130) using the similarity calculation module (140). In one embodiment, the electronic device (100) can quantify the similarity between candidate data (110) and representative data for each characteristic using the similarity calculation module (140).

[0059] In one embodiment, methods such as NDI (Normalized Difference Index), MAPE (Mean Absolute Percentage Error), JSD (Jensen-Shannon Divergence), and Cosine similarity may be used to quantify the similarity between candidate data (110) and representative data, but are not limited thereto. Refer to FIG. 9 for a specific method of calculating similarity below.

[0060] In one embodiment, the electronic device (100) can generate a matrix that quantifies the similarity between candidate data (110) and representative data according to extracted characteristics. For example, a matrix can be generated by quantifying the similarity between multiple candidate data and multiple representative data for CQI. The electronic device (100) can generate a 5x3 matrix according to characteristics such as CQI, RI, and the number of terminals for five candidate data and three representative data.

[0061] In one embodiment, the electronic device (100) can calculate similarity by applying weights to similarity values ​​between candidate data and representative data for each extracted characteristic. For example, similarity can be calculated by applying a weight of 0.2 for CQI and 0.3 for RI. Hereinafter, a specific method for calculating similarity is described with reference to FIG. 9.

[0062] The electronic device (100) can obtain additional training data (120) from candidate data (110) based on the calculated similarity and similarity range, and update information regarding the training data. The electronic device (100) can obtain additional training data (120) from candidate data (110) using a data processing module (150) and update information regarding the training data.

[0063] In one embodiment, the data processing module (150) may include a data filtering module (160) for filtering candidate data (110) and a database update module (170) for updating information regarding training data.

[0064] The electronic device (100) can select additional training data (120) from among candidate data (110) through a data filtering module (160) based on the calculated similarity and the similarity range of representative data. The similarity range refers to the range that an AI model trained using existing training data can predict, and may refer to the range of input data that the trained AI model can infer with an accuracy of a certain level or higher.

[0065] In one embodiment, the similarity range may be set differently for a plurality of representative data and for one or more characteristics. For example, among the characteristics of the first representative data, the similarity range for RI may be set to 0.4 and the similarity range for CQI may be set to 0.2, and among the characteristics of the second representative data, the similarity range for RI may be set to 0.5 and the similarity range for CQI may be set to 0.1.

[0066] In one embodiment, the electronic device (100) can determine whether to use the candidate data as training data based on whether the similarity calculated between the representative data and the candidate data through the data filtering module (160) is within the similarity range of the representative data.

[0067] In one embodiment, the electronic device (100) may not use the first candidate data as training data if, based on the calculated similarity, all similarity values ​​between the characteristics of the first candidate data included in the candidate data and the characteristics of the first representative data fall within the similarity range of the first representative data. If the first candidate data falls within the similarity range of the first representative data, which is the existing training data, the first candidate data is within the predictable range of the learned AI model, and thus the first candidate data may not be used as training data.

[0068] For example, if the similarity calculated for all characteristic values ​​including CQI, RI, number of terminals, and transmitted bytes between the first candidate data and the first representative data is within the similarity range of the first representative data, the training based on the first candidate data is within a predictable range, so the first candidate data can be filtered.

[0069] In one embodiment, the electronic device (100) may use the second candidate data as training data if, based on the calculated similarity, all similarity values ​​between the characteristics of the second candidate data included in the candidate data and the characteristics of at least one representative data do not fall within the similarity range of at least one representative data. If the second candidate data does not fall within the similarity range of at least one representative data, which is the existing training data, the second candidate data falls outside the predictable range of the learned AI model, and thus the second candidate data may be used as training data.

[0070] For example, if the similarity calculated for all characteristic values, including CQI, RI, number of terminals, transmitted bytes, etc., between the second candidate data and at least one representative data does not fall within the similarity range, the second candidate data may be identified as data that cannot be predicted by existing training data. The electronic device (100) may use the second candidate data as training data.

[0071] In one embodiment, the electronic device (100) may use the third candidate data as training data if, based on the calculated similarity, some of the similarity values ​​between the characteristics of the third candidate data and the characteristics of the first representative data are not included in the similarity range of the first representative data. If some of the characteristics of the third candidate data are not included within the similarity range of the first representative data, which is the existing training data, the third candidate data falls outside the predictable range of the learned AI model, and thus the third candidate data may be used as training data.

[0072] For example, the similarity between the third candidate data and the first representative data regarding CQI and transmitted bytes is included within the similarity range, but the similarity regarding RI and the number of terminals may not be included within the similarity range. Since some characteristics of the third candidate data fall outside the predictable range of the AI ​​model, the electronic device (100) may use the third candidate data as training data.

[0073] Refer to Fig. 5 for a specific operation for using candidate data as additional training data.

[0074] In one embodiment, the electronic device (100) can update information regarding training data through a database update module (170) based on whether the calculated similarity between representative data and candidate data is within the similarity range of the representative data. The electronic device (100) can select additional training data (120) from among candidate data (110), perform training of the AI ​​model, and then update information regarding training data based on the additional training data (120).

[0075] In one embodiment, the electronic device (100) may not use the first candidate data for updating information regarding training data, as the first candidate data is not used as training data when all similarity values ​​between the characteristics of the first candidate data included in the candidate data and the characteristics of the first representative data are included within the similarity range of the first representative data.

[0076] In one embodiment, the electronic device (100) may use the second candidate data as training data and update information regarding the training data based thereon when all similarity values ​​between the characteristics of the second candidate data included in the candidate data and the characteristics of at least one representative data do not fall within the similarity range of at least one representative data. The electronic device (100) may add the second candidate data as new representative data as the similarity values ​​between the characteristics do not fall within the similarity range.

[0077] For example, if the similarity calculated for all characteristic values ​​including CQI, RI, number of terminals, and transmitted bytes between the second candidate data and at least one representative data is not included in the similarity range, the second candidate data may be added as new representative data.

[0078] As the similarity calculated for all feature values ​​does not fall within the similarity range, the second candidate data does not possess characteristics similar to the existing training data; therefore, it can be added as new representative data and used for subsequent training data collection.

[0079] The electronic device (100) can set a similar range initial value corresponding to the addition of the second candidate data as new representative data. The electronic device (100) can set a similar range initial value so that inference beyond a preset accuracy is possible through various algorithms. For specific methods below, refer to FIG. 6.

[0080] In one embodiment, the electronic device (100) may use the third candidate data as training data and update information regarding the training data based thereon when some of the similarity values ​​between the characteristics of the third candidate data and the characteristics of the first representative data are not included in the similarity range of the first representative data. As some of the similarity values ​​between the characteristics are not included in the similarity range, the electronic device (100) may use the third candidate data to update the similarity range of the representative data.

[0081] For example, the similarity between the third candidate data and the first representative data regarding CQI and transmitted bytes may be included within the similarity range, but the similarity regarding RI and the number of terminals may not be included within the similarity range. As some characteristics of the third candidate data are not included within the similarity range of the first representative data, the similarity range of the first representative data may be extended.

[0082] Hereinafter, the operation of updating information regarding training data with candidate data is described with reference to FIGS. 5 to 7.

[0083] In one embodiment, the electronic device (100) may delete information regarding representative data as information regarding training data is updated. The electronic device (100) may delete information regarding representative data based on inclusion relationships between representative data as the similarity range of representative data is expanded.

[0084] For example, the electronic device (100) may delete information regarding the first representative data and the second representative data included in the representative data based on an update of information regarding the training data. The electronic device (100) may expand the similarity range of the second representative data based on an update of information regarding the training data. If, as a result of expanding the similarity range of the second representative data, the first representative data and the similarity range of the first representative data are included within the similarity range of the second representative data, the electronic device (100) may delete information regarding the first representative data.

[0085] For specific operations below, refer to FIG. 8.

[0086] FIG. 2 is a diagram showing a specific operation of a method for collecting training data for learning an AI model according to one embodiment.

[0087] Referring to FIG. 2, the electronic device (100) can acquire training data from candidate data for training an AI model and update information regarding the training data. The electronic device (100) can improve the learning speed of the AI ​​model by selecting only training data that cannot be predicted by the previously collected data from a vast amount of candidate data.

[0088] The electronic device (100) can acquire additional training data among candidate data based on representative data (210a, 210b, 210c) and similar ranges (220a, 220b, 220c) included in existing training data, and update information regarding training data. In the present disclosure, the similar range refers to a range that can be predicted by an AI model trained using existing training data, and may refer to a range of input data that can be inferred by the trained AI model with an accuracy of a certain level or higher.

[0089] In one embodiment of the present disclosure, the electronic device (100) may store information about at least one representative data (210a, 210b, 210c) included in existing training data. The information about the representative data (210a, 210b, 210c) may include, but is not limited to, values ​​for one or more features of the data, similarity ranges (220a, 220b, 220c), similarity range values ​​for one or more features, weight values ​​for one or more features, etc.

[0090] The electronic device (100) can extract one or more features from at least one representative data (210a, 210b, 210c) included in existing training data to obtain candidate data for training an AI model. The electronic device (100) can store values ​​for one or more features extracted from at least one representative data (210a, 210b, 210c), similarity ranges or similarity range values ​​for one or more features, weight values ​​for one or more features, etc.

[0091] In one embodiment, the electronic device (100) can acquire additional training data among candidate data (230a, 230b, 230c) based on representative data (210a, 210b, 210c) and similar ranges (220a, 220b, 220c), and update information regarding the training data.

[0092] The electronic device (100) can extract one or more features from a plurality of candidate data (230a, 230b, 230c). One or more features may include scalar features or vector features of the data. The electronic device (100) can extract one or more features from a plurality of candidate data (230a, 230b, 230c) for calculating similarity between the plurality of candidate data (230a, 230b, 230c) and representative data (210a, 210b, 210c).

[0093] The electronic device (100) can calculate the similarity between representative data (210a, 210b, 210c) and candidate data (230a, 230b, 230c) for each extracted characteristic. The electronic device (100) can quantify the similarity or difference value between the data for each characteristic.

[0094] In one embodiment, methods such as NDI (Normalized Difference Index), MAPE (Mean Absolute Percentage Error), JSD (Jensen-Shannon Divergence), and Cosine similarity may be used to quantify the similarity between candidate data (110) and representative data, but are not limited thereto. Refer to FIG. 9 for a specific method of calculating similarity below.

[0095] The electronic device (100) can obtain additional training data among candidate data (230a, 230b, 230c) based on the similarity range (220a, 220b, 220c) of the calculated similarity and representative data (210a, 210b, 210c), and update information regarding the training data.

[0096] The electronic device (100) can filter the first candidate data (230a) based on the calculated similarity if all similarity values ​​between the characteristics of the first candidate data (230a) and the characteristics of the first representative data (210a) fall within the similarity range (220a) of the first representative data (210a).

[0097] In one embodiment, the electronic device (100) may filter the first candidate data (230a) without using it as additional training data. The electronic device (100) may determine that the first candidate data (230a) is within a predictable range by learning from the first representative data (210a), as all characteristics of the first candidate data (230a) are within the similarity range (220a) of the first representative data (210a).

[0098] In one embodiment, the electronic device (100) may filter out data within a predictable range from a vast amount of data and not use it as training data. The electronic device (100) may filter out first candidate data (230a) within a predictable range and not use it as additional training data.

[0099] The electronic device (100) may use the second candidate data (210b) as additional training data if, based on the calculated similarity, all similarity values ​​between the characteristics of the second candidate data (230b) and the characteristics of the representative data (210a, 210b, 210c) are not included within the similarity range (220a, 220b, 220c) of the representative data (210a, 210b, 210c).

[0100] In one embodiment, the electronic device (100) may determine that the second candidate data (230b) is not within the range predictable by learning from the representative data (210a, 210b, 210c) because all characteristics of the second candidate data (230b) are not included within the similarity range (220a, 220b, 220c) of the representative data (210a, 210b, 210c).

[0101] In one embodiment, the electronic device (100) may determine the second candidate data (230b) as additional training data and use it for training an AI model. The electronic device (100) may use the second candidate data (230b) for training an AI model and, based thereon, update information regarding the training data. The electronic device (100) may add the second candidate data (230b) as new representative data (240).

[0102] In one embodiment, the electronic device (100) may use the second candidate data (230b) as new representative data because it is not included within all similar ranges (220a, 220b, 220c) as it does not have characteristics similar to existing representative data (210a, 210b, 210c). The electronic device (100) may add the second candidate data (230b) as new representative data and subsequently use it in a training data collection processor.

[0103] The electronic device (100) can set a corresponding similarity range initial value by adding the second candidate data (230b) as new representative data. The electronic device (100) can set a similarity range initial value so that inference beyond a preset accuracy is possible through various algorithms. For specific methods below, refer to FIG. 6.

[0104] The electronic device (100) may use the third candidate data (210c) as additional training data if, based on the calculated similarity, some of the similarity values ​​between the features of the third candidate data (230c) and the features of the second representative data (210b) are not included within the similarity range (220b) of the second representative data (210b).

[0105] In one embodiment, the electronic device (100) may determine that the third candidate data (230c) is not within a range of some predictable ranges by learning from the second representative data (210b), as some characteristics of the third candidate data (230c) are not included within the similar range (220b) of the second representative data (210b).

[0106] In one embodiment, the electronic device (100) may determine the third candidate data (230c) as additional training data and use it for training an AI model. The electronic device (100) may use the third candidate data (230c) for training an AI model and, based thereon, update information regarding the training data. The electronic device (100) may update the similarity range (220b) of the second representative data (210b) based on the third candidate data (230c).

[0107] In one embodiment, the electronic device (100) can update (250) the similarity range (220b) of the second representative data (210b) to include the characteristic that is least outside the similarity range among the characteristics of the third candidate data (230c) as some characteristics of the second representative data (210b) are not included within the similarity range (220b) of the second representative data (210b).

[0108] Hereinafter, the operation to update information regarding specific training data is described with reference to FIGS. 5 to 7.

[0109] The electronic device (100) may delete information regarding representative data as it updates information regarding training data. In one embodiment, the electronic device (100) may delete information regarding representative data based on inclusion relationships between representative data as the similarity range of representative data is updated (250).

[0110] For example, the electronic device (100) may expand the similarity range (220b) of the second representative data (210b) based on an update of information regarding training data. If, following the expansion (250) of the similarity range of the second representative data, the fourth representative data (not shown) and the similarity range of the fourth representative data are included within the similarity range (220b) of the second representative data, the electronic device (100) may delete information regarding the fourth representative data.

[0111] For specific deletion operations below, refer to FIG. 8.

[0112] FIG. 3 is a diagram illustrating the operation of collecting training data for learning an AI model according to one embodiment.

[0113] Referring to FIG. 3, this is a diagram of a flow for collecting training data for training an AI model, illustrating a method of operation that adds AI-specific filtering and an AI-specific database in addition to the existing data collection operation.

[0114] The flow for collecting existing data is as follows. A vast amount of data collected from multiple base stations is transmitted to the AI ​​model for training. Hereinafter, the data collected from multiple base stations is referred to as candidate data.

[0115] Multiple base stations can collect candidate data including network-related data, traffic data, event-based data, etc. In one embodiment, multiple base stations can collect data by monitoring data in real time or through sampling based on preset time intervals.

[0116] A USM (User Service Module) can perform a preprocessing process by integrating data received from multiple base stations. In one embodiment, the USM can perform a preprocessing process on data received from multiple base stations by removing duplicate data and filtering it. In one embodiment, the USM can transform data collected from multiple base stations and structure it into a format suitable for an AI model.

[0117] Specifically, the data collection process for an AI model may include data cleaning, data filtering, data transformation, and data integration processes. Data cleaning is a process of removing errors from data and improving quality, and may include the removal of duplicate data. The data filtering process refers to selecting only data suitable for the learning purpose. The data transformation process refers to converting data into a form suitable for modeling, and in this disclosure, may refer to the process of converting data to be suitable for an AI model. The data integration process may refer to integrating data from multiple sources into a single data set to create a consistent dataset.

[0118] Collected data can be stored in a database. The collected data can be used to train AI models. In the case of existing data collection flows, there are limitations in that applying vast amounts of data directly to AI models results in inefficient storage space utilization and inefficient training of the AI ​​models.

[0119] In the present disclosure, to efficiently utilize storage space and increase the learning efficiency of the AI ​​model, data for training the AI ​​model is collected by adding an AI-dedicated filtering (310) and an AI-dedicated database (320).

[0120] AI-specific filtering (310) performs the role of filtering input values ​​that the trained AI model can predict based on existing training data used for training the AI ​​model. In one embodiment, AI-specific filtering (310) can select whether to use candidate data as training data based on representative data included in the existing training data and the range of similarity of the representative data.

[0121] In one embodiment, the AI-only filtering (310) can filter the candidate data without using it for training the AI ​​model when the candidate data obtained from a plurality of base stations falls within the similar range of the representative data, as the candidate data is within the range predictable by the AI ​​model trained based on the representative data.

[0122] In one embodiment, AI-only filtering (310) can be used as training data for learning an AI model without filtering candidate data when candidate data is not included within the similar range of representative data, as the candidate data is outside the range predictable by the AI ​​model learned based on representative data.

[0123] For specific operation methods below, refer to the descriptions in other drawings.

[0124] The AI ​​dedicated database (320) can store information about representative data that represents the ranges of training data among the data used for training the AI ​​model, and information about the similar ranges of the representative data.

[0125] In one embodiment, the AI-dedicated database (320) may store information regarding representative data, and may store one or more characteristics of the representative data. The AI-dedicated database (320) may store similarity ranges for one or more characteristics of representative data. For example, the AI-dedicated database (320) may store similarity ranges for one or more characteristic values ​​of multiple representative data and similarity ranges for one or more characteristics, such as a similarity range for the CQI of the first representative data, a similarity range for the RI of the first representative data, a similarity range for the CQI of the second representative data, and a similarity range for the RI of the second representative data.

[0126] The present disclosure describes a method for enabling an electronic device (100) to efficiently perform learning of an AI model by adding an AI-dedicated filtering (310) and an AI-dedicated database (320) to an existing data collection flow.

[0127] FIG. 4 is a flowchart illustrating the operation of collecting training data for learning an AI model according to one embodiment.

[0128] Referring to FIG. 4, the electronic device (100) can acquire training data for learning an AI model and perform training of the AI ​​model based thereon. The electronic device (100) can acquire additional training data from a plurality of candidate data and update information regarding the training data.

[0129] In step S410, the electronic device (100) can acquire multiple candidate data that can be used to train an AI model. The electronic device (100) can acquire multiple candidate data generated in various network environments for training an AI model in a wireless communication system.

[0130] The electronic device (100) can acquire multiple candidate data for training an AI model. The electronic device (100) can train the AI ​​model by acquiring additional training data from the multiple candidate data through subsequent steps. In one embodiment, the electronic device (100) collecting data in a wireless communication system can acquire data from multiple base stations.

[0131] In step S420, the electronic device (100) can extract one or more characteristics for each of the multiple candidate data. The electronic device (100) can extract one or more characteristics representing the attributes of the data according to the type of the multiple candidate data.

[0132] In one embodiment, the electronic device (100) can extract one or more characteristics from a plurality of candidate data obtained from a wireless communication system. One or more characteristics may include scalar characteristics or vector characteristics.

[0133] For example, scalar characteristics regarding network data may include, but are not limited to, the number of terminals, transmitted bytes, received bytes, etc. Vector characteristics regarding network data may include, but are not limited to, CQI, which represents quality information of the wireless channel, RI, which represents the number of spatial streams in the antenna system.

[0134] In one embodiment, vector characteristics are extracted to have the same vector size for a plurality of candidate data and representative data.

[0135] In step S430, the electronic device (100) can calculate the similarity between one or more features of at least one representative data included in the existing training data used to train the AI ​​model and one or more features extracted for each of the multiple candidate data.

[0136] The electronic device (100) can extract one or more features from at least one representative data included in the existing training data used to train the AI ​​model. One or more features extracted from the representative data may correspond to one or more features extracted from the candidate data extracted in step S420.

[0137] In one embodiment, the electronic device (100) may load and use one or more characteristics of at least one representative data stored in a database. The electronic device (100) may store one or more characteristics extracted from at least one representative data in a database prior to acquiring a plurality of candidate data, and may use them in a subsequent procedure.

[0138] The electronic device (100) can calculate the similarity between at least one representative data and a plurality of candidate data. Similarity is a standard for quantitatively measuring the degree of correlation or agreement between data, and may refer to a value that quantifies the difference according to one or more characteristics. In one embodiment, similarity may refer to a matrix that quantifies the difference according to one or more characteristics between a plurality of candidate data and a plurality of representative data. For specific examples below, refer to FIG. 9.

[0139] In one embodiment, the electronic device (100) can calculate the similarity between at least one representative data and a plurality of candidate data for each of one or more characteristics. The electronic device (100) may use NDI, MAPE, JSD, Cosine similarity calculation methods, etc., to calculate the similarity between the representative data and the candidate data, but is not limited thereto.

[0140] In one embodiment, the electronic device (100) can calculate similarity by applying weights to similarity values ​​between candidate data and representative data for one or more characteristics. Hereinafter, a specific method for calculating similarity is referenced in FIG. 9.

[0141] In step S440, the electronic device (100) can acquire additional training data among a plurality of candidate data based on the calculated similarity and the similarity range of at least one representative data. The electronic device (100) can determine that among the plurality of candidate data, data that cannot be predicted based on existing training data is additional training data.

[0142] The electronic device (100) can select additional training data based on the similarity and the similarity range of representative data calculated in step S430. The similarity range refers to the range that an AI model trained using existing training data can predict, and may refer to the range of input data that the trained AI model can infer with a certain level of accuracy.

[0143] In one embodiment, the similarity range may be set differently for at least one characteristic for at least one representative data. For example, among the characteristics of the first representative data, the similarity range for RI may be set to 0.4 and the similarity range for CQI may be set to 0.2, and among the characteristics of the second representative data, the similarity range for RI may be set to 0.5 and the similarity range for CQI may be set to 0.1.

[0144] In one embodiment, the electronic device (100) may determine whether to use additional training data based on whether the similarity calculated between at least one representative data and a plurality of candidate data is within the similarity range of the representative data. The electronic device (100) may determine whether to use additional training data based on whether the candidate data is within the similarity range of the representative data, that is, whether the candidate data can be inferred from the representative data.

[0145] In one embodiment, the electronic device (100) may filter the first candidate data without using it as training data if, based on the similarity calculated in step S430, all similarity values ​​between one or more characteristics of the first candidate data included in the plurality of candidate data and one or more characteristics of the first representative data fall within the similarity range of the first representative data.

[0146] If all similarity values ​​for one or more features between the first candidate data and the first representative data are within the similarity range of the first representative data, the first candidate data is the predictable range of the AI ​​model learned by the first representative data, so the electronic device (100) can filter the first candidate data without using it as training data.

[0147] In one embodiment, the electronic device (100) may use the first candidate data as additional training data if, based on the similarity calculated in step S430, all similarity values ​​between one or more characteristics of the second candidate data included in the plurality of candidate data and one or more characteristics of at least one representative data do not fall within the similarity range of at least one representative data.

[0148] If all of the similarity values ​​for one or more features between the second candidate data and at least one representative data are not included within the similarity range of the at least one representative data, the second candidate data is outside the predictable range of the AI ​​model learned by the at least one representative data, and the electronic device (100) can use the second candidate data as additional training data.

[0149] In one embodiment, the electronic device (100) may use the third candidate data as additional training data if, based on the similarity calculated in step S430, a portion of the similarity values ​​between one or more characteristics of the third candidate data included in the plurality of candidate data and one or more characteristics of the first representative data falls within the similarity range of the first representative data.

[0150] If some of the similarity values ​​for one or more features between the third candidate data and the first representative data do not fall within the similarity range of the first representative data, the third candidate data falls outside the predictable range of the AI ​​model learned from the first representative data, and the electronic device (100) can use the third candidate data as additional training data.

[0151] For specific operations below, refer to FIG. 5.

[0152] In step S450, the electronic device (100) can train an AI model using additional training data. The electronic device (100) can train an AI model using candidate data that falls outside the range predictable by the AI ​​model trained in step S440 as additional training data.

[0153] The electronic device (100) can acquire only the data that falls outside the range predictable by the learned AI model from among a vast amount of candidate data as additional training data in step S440. The electronic device (100) can perform efficient training of the AI ​​model by training the AI ​​model based on the acquired additional training data.

[0154] In step S460, the electronic device (100) can update information regarding training data based on additional training data. Based on the acquired additional training data, the electronic device (100) can update information regarding training data to collect training data for learning the AI ​​model thereafter.

[0155] The electronic device (100) can update information regarding the training data to reflect the predictable range of the AI ​​model that has changed as the AI ​​model is trained with additional training data.

[0156] In one embodiment, the electronic device (100) filters the first candidate data included in the plurality of candidate data within a predictable range based on the first representative data, so that the first candidate data may not be used for updating information regarding the training data.

[0157] In one embodiment, the electronic device (100) can update information regarding training data based on the second candidate data, such that all similarity values ​​between one or more characteristics of the second candidate data included in the plurality of candidate data and one or more characteristics of at least one representative data are not included within the similarity range of at least one representative data.

[0158] The electronic device (100) may add a second candidate data as new representative data as all similarity values ​​between the characteristics are not included within the similarity range. For example, if the similarity calculated for all characteristic values ​​including CQI, RI, number of terminals, transmitted bytes, etc. between the second candidate data and the first representative data is not included within the similarity range, the second candidate data may be added as new representative data.

[0159] As the similarity calculated for all feature values ​​does not fall within the similarity range, the second candidate data does not possess characteristics similar to the existing training data; therefore, it can be added as new representative data and used for subsequent training data collection.

[0160] The electronic device (100) can set a similar range initial value corresponding to the addition of the second candidate data as new representative data. The electronic device (100) can set a similar range initial value so that inference beyond a preset accuracy is possible through various algorithms. For specific methods below, refer to FIG. 6.

[0161] In one embodiment, the electronic device (100) can update information regarding training data based on the third candidate data, as some of the similarity values ​​between one or more characteristics of the third candidate data included in the plurality of candidate data and one or more characteristics of the first representative data are not included within the similarity range of the first representative data.

[0162] The electronic device (100) can update the similarity range of the first representative data based on the third candidate data as some of the similarity values ​​between the characteristics are not included within the similarity range of the first representative data. For example, the similarity between the third candidate data and the first representative data for CQI and transmission bytes may be included within the similarity range, but the similarity for RI and the number of terminals may not be included within the similarity range. As some characteristics of the third candidate data are not included within the similarity range of the first representative data, the similarity range of the first representative data may be extended.

[0163] The electronic device (100) can extend the similarity range of the first representative data to include the characteristic of the third candidate data that deviates least from the similarity range, for a characteristic of the third candidate data that does not fall within the similarity range of the first representative data among the characteristics of the third candidate data. For specific operations below, refer to FIGS. 5 to 7.

[0164] In one embodiment, the electronic device (100) may delete information regarding at least one representative data as information regarding training data is updated. The electronic device (100) may delete information regarding the representative data based on the inclusion relationship between the representative data as the similarity range of at least one representative data is expanded.

[0165] For example, as the electronic device (100) expands the similarity range regarding the first representative data, if the data in the second representative and the similarity range of the second representative data are included within the similarity range of the first representative data, the information regarding the second representative data may be deleted. Since the second representative data and the similarity range of the second representative data being included within the similarity range of the first representative data means that the predictable range from the second representative data is included within the predictable range from the first representative data, the electronic device (100) may delete the information regarding the second representative data.

[0166] For specific operations below, refer to FIG. 9.

[0167] FIG. 5 is a diagram illustrating an operation for processing data based on similarity according to one embodiment.

[0168] Referring to FIG. 5, the electronic device (100) can determine whether to use additional training data by calculating the similarity between candidate data and representative data, and can update information regarding the training data.

[0169] The electronic device (100) can calculate a similarity between one or more characteristics of candidate data and one or more characteristics of representative data. In the present disclosure, similarity may mean a value that quantifies the difference between one or more characteristics as a criterion for quantitatively measuring the relationship between data.

[0170] The electronic device (100) may use NDI, MAPE, JSD, Cosine similarity calculation methods, etc., to calculate similarity for one or more characteristics between candidate data and representative data, but is not limited thereto. For specific similarity calculation methods below, refer to FIG. 9.

[0171] In one embodiment, the electronic device (100) can determine whether to use candidate data as additional training data based on the calculated similarity and the similarity range of representative data, and can update information regarding the training data. The electronic device (100) can increase the learning efficiency of the AI ​​model by selecting training data based on similarity and the similarity range, thereby not storing all of the vast amount of data and not using it for learning.

[0172] In one embodiment, the electronic device (100) can filter the candidate data (530) when all of the similarity values ​​between one or more characteristics of the candidate data and the representative data are included within the similarity range (520) of the representative data.

[0173] For example, the electronic device (100) can filter the first candidate data based on the calculated similarity if all similarity values ​​between one or more characteristics of the first candidate data and one or more characteristics of the first representative data fall within the similarity range of the first representative data.

[0174] If all of the similarity values ​​between one or more features for the first candidate data and the first representative data are within the similarity range of the first representative data, the electronic device (100) can determine the first candidate data as an input value within the range that can be inferred by an AI model learned using the first representative data.

[0175] If the first candidate data falls within the similarity range of the first representative data, which is the existing training data, then the first candidate data is within the predictable range of the learned AI model, so the first candidate data can be filtered out without being used as training data.

[0176] The electronic device (100) is not used to update information regarding training data, as it is not used for training the AI ​​model with respect to data among the candidate data that is not used as additional training data.

[0177] In one embodiment, if the electronic device (100) finds that all of the similarity values ​​between one or more features of the candidate data and the representative data are not included within the similarity range of the representative data (540), the candidate data can be used as additional training data and added as new representative data (550).

[0178] For example, the electronic device (100) may use the second candidate data as additional training data if, based on the calculated similarity, all similarity values ​​between one or more features of the second candidate data and one or more features of at least one representative data do not fall within the similarity range of at least one representative data.

[0179] If all of the similarity values ​​between one or more features for the second candidate data and at least one representative data are not included within the similarity range of at least one representative data, the electronic device (100) may determine the second candidate data as an input value that is not within the range that can be inferred by an AI model learned using at least one representative data.

[0180] For example, at least one representative data may include the first to fourth representative data. For the second candidate data, the similarity value with respect to one or more features with respect to the first to fourth representative data may not be included within the similarity range of the first to fourth representative data, respectively. The electronic device (100) may determine that the second candidate data is outside the predictable range of the AI ​​model learned by the first to fourth representative data, and may identify / determine / decide the second candidate data as additional training data for learning the AI ​​model.

[0181] The electronic device (100) can update information regarding training data based on the second candidate data, as the second candidate data falls outside the range that can be inferred by at least one representative data, which is existing training data. In one embodiment, the electronic device (100) can add the second candidate data as new representative data, as at least one of the characteristics of the second candidate data is not similar to all of the at least one representative data.

[0182] The electronic device (100) can set a similar range initial value corresponding to the addition of the second candidate data as new representative data. The electronic device (100) can set a similar range initial value so that inference beyond a preset accuracy is possible through various algorithms. For specific methods below, refer to FIG. 6.

[0183] The electronic device (100) can proceed with a training data collection procedure for learning the AI ​​model based on information regarding representative data updated to include second candidate data.

[0184] In one embodiment, the electronic device (100) can use the candidate data as additional training data and update the similarity range of at least one representative data (570) when some of the similarity values ​​between one or more features of the candidate data and the representative data are not included within the similarity range of the representative data (560).

[0185] For example, the electronic device (100) may use the second candidate data as additional training data if, based on the calculated similarity, some of the similarity values ​​between one or more features of the third candidate data and one or more features of the first representative data do not fall within the similarity range of the first representative data.

[0186] If some of the similarity values ​​between one or more features for the third candidate data and the first representative data are not included within the similarity range of the first representative data, the electronic device (100) may determine that the third candidate data is an input value that is not within the range that can be inferred by an AI model learned using the first representative data.

[0187] The electronic device (100) can update information regarding training data based on the third candidate data, as the third candidate data falls outside the range that can be inferred by at least one representative data, which is existing training data. In one embodiment, the electronic device (100) can update the similarity range of the first representative data based on the third candidate data, as one or more characteristics of the third candidate data are not somewhat similar to the first representative data.

[0188] In one embodiment, the electronic device (100) may update the similarity range of the first representative data based on one or more characteristics of the third candidate data that fall outside the similarity range of the first representative data. The electronic device (100) may extend the similarity range of the first representative data to include, for one or more characteristics of the third candidate data, the characteristic that falls outside the similarity range of the first representative data the least among the characteristics that fall outside the similarity range of the first representative data.

[0189] For example, among one or more characteristics of the third candidate data, the CQI and the number of terminals may deviate from the similarity range of the first representative data. In the case where the similarity between the third candidate data and the first representative data for each is 0.45 and 0.38, and the similarity range of the first data is 0.4 and 0.3, the CQI deviates less from the similarity range than the number of terminals. The electronic device (100) can update the similarity range to 0.45 so that the similarity range of the first data can include the CQI value of the third candidate data.

[0190] The electronic device (100) can proceed with a training data collection procedure for learning the AI ​​model based on information including a similar range of representative data updated based on third candidate data.

[0191] FIG. 6 is a diagram illustrating the operation of adding representative data as an example of updating information regarding training data according to one embodiment.

[0192] Referring to FIG. 6, the electronic device (100) can use candidate data that is unpredictable with existing training data for training an AI model and update information regarding the training data based on this.

[0193] The electronic device (100) can determine whether to use multiple candidate data for training an AI model based on representative data (610) included in existing training data and the similarity range (620) of the representative data. The electronic device (100) can select unpredictable data from existing trained AI models from among a vast amount of candidate data to use as training data for efficient training of the AI ​​model.

[0194] In one embodiment, the electronic device (100) can calculate a similarity value between one or more features of representative data (610) and candidate data (630) and determine whether to use the candidate data (630) for training an AI model based on the similarity range (620) of the representative data.

[0195] If all of the similarity values ​​between one or more features of representative data (610) and candidate data (630) are not included within the similarity range (620) of the representative data, the electronic device (100) determines the candidate data (630) as additional training data and can use it for training an AI model.

[0196] The electronic device (100) determines that the candidate data (630) is outside the predictable range of the AI ​​model learned by the representative data (610) because all of the similarity values ​​between one or more features are not included in the similarity range (620) of the representative data, and can use the candidate data (630) for training the AI ​​model.

[0197] The electronic device (100) can add candidate data (630) as new representative data (640) as all similarity values ​​between one or more features are not included in the similarity range (620) of the representative data. The electronic device (100) can perform a training data collection procedure for training the AI ​​model based on the updated training data information as it adds the new representative data.

[0198] For example, the electronic device (100) may add the candidate data (630) as new representative data (640) if the similarity value calculated for all characteristics including CQI, RI, number of terminals, and transmitted bytes between the candidate data (630) and the representative data (610) falls outside the similarity range (620) of the representative data for each characteristic.

[0199] The electronic device (100) can initially set a similarity range for the new representative data by adding candidate data (630) as new representative data (640). In one embodiment, the electronic device (100) can set a similarity range for the new representative data based on a preset accuracy. For example, the electronic device (100) can set an initial similarity range by adjusting the similarity range for one or more characteristics of the new representative data so that the evaluation result through a test set is greater than or equal to a preset accuracy.

[0200] In one embodiment, the electronic device (100) can set an initial value for a similar range for new representative data based on the MSD (Mean Squared Deviation) method. By calculating statistics on the similar range of existing representative data and identifying the characteristics of the similar range distribution, an initial value for a similar range for new representative data can be set. By calculating the mean squared difference between existing data and new data, if the MSD value is small, the existing data is set as the initial value, and if the MSD value is large, the initial value for new representative data can be set to nothing or set through an additional search process.

[0201] The electronic device (100) is not limited to the above algorithm and can set an initial value for the similarity range of new representative data using various algorithms, and can modify / set the initial value for the similarity range so that inference greater than the pre-set accuracy is possible through a test procedure for the AI ​​model.

[0202] FIG. 7 is a diagram illustrating an operation to update a similar range as an example of an information update regarding training data according to one embodiment.

[0203] Referring to FIG. 7, the electronic device (100) can use candidate data that is unpredictable with existing training data for training an AI model and update information regarding the training data based on this.

[0204] The electronic device (100) can determine whether to use multiple candidate data for training an AI model based on representative data (710) included in existing training data and the similarity range (720) of the representative data. The electronic device (100) can select unpredictable data from existing trained AI models from among a vast amount of candidate data to use as training data for efficient training of the AI ​​model.

[0205] In one embodiment, the electronic device (100) can calculate a similarity value between one or more features of representative data (710) and candidate data (730) and determine whether to use candidate data (730) for training an AI model based on the similarity range (720) of the representative data.

[0206] If some of the similarity values ​​between one or more features of representative data (710) and candidate data (730) are not included within the similarity range (720) of the representative data, the electronic device (100) determines the candidate data (730) as additional training data and can use it for training an AI model.

[0207] The electronic device (100) determines that the candidate data (730) is outside the predictable range of the AI ​​model learned by the representative data (710) as some of the similarity values ​​between one or more features are not included in the similarity range (720) of the representative data, and can use the candidate data (730) for training the AI ​​model.

[0208] The electronic device (100) can update (740) the similarity range (720) of representative data based on candidate data (730) as some of the similarity values ​​between one or more features are not included in the similarity range (720) of representative data. The electronic device (100) can perform a training data collection procedure for training an AI model based on information regarding training data including the updated similarity range of representative data.

[0209] For example, if the similarity value calculated for CQI and the number of terminals among the characteristics including CQI, RI, number of terminals, and transmitted bytes between candidate data (730) and representative data (710) falls outside the similarity range (720) of representative data for each characteristic, the electronic device (100) can extend the similarity range (720) of representative data based on candidate data (730).

[0210] In one embodiment, the electronic device (100) can extend (740) the similarity range (720) of the representative data (710) to include the characteristic that deviates least from the similarity range (720) of the representative data (710) among at least one characteristic of the candidate data (730) that deviates from the similarity range (720) of the representative data (710), based on the result of calculating similarity by one or more characteristics between the candidate data (730) and the representative data (710).

[0211] For example, among one or more features of candidate data (730), the number of RIs and terminals may be outside the similarity range (720) of representative data. If the similarity between candidate data (730) and representative data (710) for each feature is 0.34 and 0.41, and the similarity range (720) of representative data for each feature is 0.3 and 0.4, the electronic device (100) can update the similarity range of representative data for the number of terminals that are less outside the similarity range of the two features.

[0212] FIG. 8 is a diagram illustrating the operation of deleting representative data according to an information update regarding training data according to one embodiment.

[0213] Referring to FIG. 8, the electronic zi (100) can use candidate data in an AI model that is unpredictable with existing training data, and update information regarding the training data based on this.

[0214] The electronic device (100) determines candidate data as additional training data to perform training of the AI ​​model, and can delete information regarding some representative data according to the update of information regarding training data for effective operation of the database.

[0215] The electronic device (100) can collect training data for learning an AI model based on a first representative data (810a) and a similarity range (820a) of the first representative data and a second representative data (810b) and a similarity range (820b) of the second representative data. The electronic device (100) can determine whether to use candidate data (830) as training data for learning an AI model by calculating the similarity with the representative data.

[0216] In one embodiment, the electronic device (100) may update the similarity range (820a) of the first representative data based on the candidate data (830) when some of the similarity values ​​between one or more characteristics of the candidate data (830) and the first representative data (810a) are not included within the similarity range (820a) of the first representative data. For specific operations below, refer to the description of FIG. 7.

[0217] The electronic device (100) may use candidate data (830) for training an AI model and, based thereon, perform a similarity range update (840) of the first representative data. In one embodiment, according to the similarity range update (840) of the first representative data, the second representative data (810b) and the similarity range (820b) of the second representative data may be included within the extended similarity range (850) of the first representative data.

[0218] The electronic device (100) can delete information regarding at least one representative data as information regarding training data is updated. The electronic device (100) can delete information regarding the representative data based on the inclusion relationship between the representative data as the similarity range of at least one representative data is expanded.

[0219] In one embodiment, the electronic device (100) may delete information regarding the second representative data as the second representative data (810b) and the similarity range (820b) of the second representative data are included within the extended similarity range (850) of the first representative data. Since the second representative data (810b) and the similarity range (820b) of the second representative data are included within the extended similarity range (9850) of the first representative data, the range predictable from the second representative data (810b) is included within the range predictable from the first representative data (810a), the electronic device (100) may delete information regarding the second representative data for efficient database operation.

[0220] FIG. 9 is a diagram relating to example data for specifically explaining the operation of collecting training data for learning an AI model according to one embodiment.

[0221] Referring to Fig. 9, we intend to explain how to determine whether to use candidate data and representative data for training an AI model as training data, and how to update information regarding the training data based on this.

[0222] Information regarding existing training data may include two representative data and values ​​for a total of four characteristics. An electronic device (100) can extract CQI, RI, the number of terminals, and transmitted data bytes as characteristics from data regarding a network in a wireless communication system. In one embodiment, vector characteristics such as CQI and RI may be extracted in vector form, and scalar characteristics such as the number of terminals and transmitted data bytes may be extracted in scalar form. FIG. 9 illustrates a table (910) for characteristic values ​​extracted for representative data.

[0223] An electronic device (100) can acquire candidate data for training an AI model. In one embodiment, the electronic device (100) can acquire three candidate data. The electronic device (100) can extract four features from the acquired candidate data to correspond to representative data. FIG. 9 illustrates a table (920) for feature values ​​extracted from the candidate data.

[0224] In one embodiment, the electronic device (100) can calculate similarity based on four characteristics extracted for two representative data and three candidate data. Similarity is a value that quantifies the difference between characteristics, and similarity between representative data and candidate data can be calculated through various algorithms.

[0225] In one embodiment, the electronic device (100) can generate a matrix by quantifying the difference between representative data and candidate data for each characteristic. The electronic device (100) can quantify the difference for each characteristic using an algorithm for scalar characteristics and vector characteristics, respectively.

[0226] For scalar characteristics, similarity can be calculated using algorithms such as NDI (Normalized Difference Index) and MAPE (Mean Absolute Percentage Error), but is not limited to these.

[0227] NDI is an indicator used to normalize the difference between two values ​​for comparison, and is mainly used to evaluate the relative difference between values. The difference between x and y can be normalized using the following mathematical formula (1).

[0228] ... mathematical formula (1)

[0229] The DI value is between 0 and 1, and a smaller NDI value indicates a higher degree of similarity between x and y.

[0230] MAPE is a metric primarily used to evaluate prediction performance, and can normalize the difference between x and y through the following mathematical formula (2).

[0231] ... mathematical formula (2)

[0232] The MAPE value is between 0 and 1, and a smaller MAPE value indicates a higher degree of similarity between x and y.

[0233] For vector characteristics, similarity can be calculated using methods such as JSD (Jensen-Shannon Divergence) and Cosine similarity calculation, but is not limited thereto.

[0234] JSD is an indicator that measures the difference between two vectors, and the difference between two vectors p and q can be normalized through the following mathematical formula (3).

[0235] ... mathematical formula (3)

[0236] The JSD value is between 0 and 1, where a value of 0 means the two vectors are identical and a value of 1 means the two vectors are completely different.

[0237] Cosine similarity measures similarity using the cosine value of the angle between two vectors, and the difference between two vectors p and q can be normalized through the following mathematical formula (4).

[0238] ... mathematical formula (4)

[0239] The cosine similarity value has a value between -1 and 1, where a value of '1' means the two vectors are highly similar, '0' means the two vectors are unrelated, and '-1' means the two vectors are opposite. To match the cosine similarity value with other algorithm results, the electronic device (100) can process the similarity value through additional processing so that it has a value between 0 and 1, and has a value of '0' if the two vectors are identical and a value of '1' if the two vectors are completely different.

[0240] The electronic device (100) can generate a matrix that quantifies the difference values ​​for representative data and candidate data for each of the four features. For example, for a specific feature, a matrix as follows can be generated.

[0241] ... mathematical formula (5)

[0242] According to the table shown in Fig. 9, the matrix quantifying the difference between RI and the number of terminals (UE) among the four characteristics is as follows.

[0243] ...mathematical formula (6)

[0244] ... mathematical formula (7)

[0245] In one embodiment, the electronic device (100) can quantify the similarity between representative data and candidate data for each of the four characteristics, or generate an integrated matrix by applying weights to each characteristic. For example, a matrix can be generated by applying weights to each characteristic as in the following mathematical formula (8).

[0246] ... mathematical formula (8)

[0247] In one embodiment, the electronic device (100) can select training data among candidate data based on the calculated similarity and the similarity range of representative data. The similarity range refers to a range that can be predicted by an AI model trained using existing training data, and may refer to a range of input data that the trained AI model can infer with a certain level of accuracy.

[0248] In one embodiment, the similarity range of representative data can be set by characteristic. For example, the similarity range for RI for two representative data is As for the UE, the similar range is It can be set to.

[0249] Based on a matrix that quantifies the difference in the number of RIs and terminals for two representative data and three candidate data, and the similarity range of the representative data, the electronic device (100) can determine whether to use the candidate data as training data.

[0250] For each column of the matrix quantifying differences by characteristic, '0' can be converted if it is within the similarity range, and '1' if it is not included in the similarity range. For example, the matrix can be transformed as shown in the following mathematical formulas (9) and (10).

[0251] ... mathematical formula (9)

[0252] ... mathematical formula (10)

[0253] The electronic device (100) can generate a final difference quantification matrix by summing the matrices transformed by characteristic as above. In the case based on only two characteristics, the final difference quantification matrix can be generated as follows.

[0254] ... mathematical formula (11)

[0255] In one embodiment, the electronic device (100) may filter out candidate data and not use it as training data if there is a '0' in one element for each row. The electronic device (100) may filter out candidate data and not use it as training data if it determines that candidate data corresponding to an element is included within the similar range of representative data if there is a '0' in one element for each row.

[0256] In one embodiment, the electronic device (100) may select and use the candidate data as additional training data for learning an AI model if the element in each row does not contain '0'. The electronic device (100) may update information regarding the training data based on the additional training data. Specific operations are described below with reference to FIGS. 1 to 8.

[0257] FIG. 10 is a schematic block diagram of an electronic device according to one embodiment.

[0258] Referring to FIG. 10, the electronic device (100) may include memory (1010), a processor (1020), and a communication unit (1030). However, not all of the illustrated components are essential components of the electronic device (100). The electronic device (100) may be implemented by more components than those illustrated, or by fewer components.

[0259] The memory (1010) can store a program for processing and controlling the processor (1020) and can store data that is input to or output from the electronic device (100). The memory (1010) may include at least one type of storage medium among flash memory type, hard disk type, multimedia card micro type, card type memory (e.g., SD or XD memory), RAM (Random Access Memory), SRAM (Static Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk, and optical disk. In one embodiment, the memory (1010) may include a weight memory for storing weights and a feature vector memory for including feature vectors which are intermediate output values ​​of a neural network model, but is not limited thereto.

[0260] In one embodiment of the present disclosure, the memory (1010) can store candidate data for training an AI model by the operation of the processor (1020). The memory (1010) can store information regarding one or more features extracted from the candidate data by the operation of the processor (1020).

[0261] In one embodiment, the memory (1010) can store information about existing training data for learning an AI model by the operation of the processor (1020). The memory (1010) can store information about representative data, information about one or more characteristics of representative data, and information about the similarity range of representative data.

[0262] In one embodiment, the memory (1010) can store similarity for one or more characteristics of representative data and candidate data by the operation of the processor (1020). The memory (1010) can store similarity calculated for one or more characteristics.

[0263] In one embodiment, the memory (1010) can update and store information regarding training data by the operation of the processor (1020). The memory (1010) can store new representative data or updated information of a similar range of representative data according to the update of information regarding training data.

[0264] The processor (1020) controls the overall operation of the electronic device (100). For example, the processor (1020) can perform the functions of the electronic device (100) described in the present disclosure by executing one or more instructions stored in memory (1010). In this case, memory (1010) may store one or more instructions that can be executed by the processor (1020). Additionally, the processor (1020) can store one or more instructions in internally provided memory and control the execution of the one or more instructions stored in the internally provided memory so that the aforementioned operations are performed. That is, the processor (1020) can perform a predetermined operation by executing at least one instruction or program stored in internal memory or memory (1010) provided within the processor (1020).

[0265] The processor (1020) may include one or more processors. In this case, the one or more processors may be general-purpose processors such as CPUs, APs, DSPs (Digital Signal Processors), graphics-dedicated processors such as GPUs, VPUs (Vision Processing Units), or artificial intelligence-dedicated processors such as NPUs. If one or more processors are artificial intelligence-dedicated processors, the artificial intelligence-dedicated processors may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0266] In one embodiment of the present disclosure, the processor (1020) may receive candidate data that can be used for training an AI model through a communication unit (1030). The processor (1020) may extract one or more features from the candidate data.

[0267] In one embodiment, the processor (1020) can extract one or more features from representative data. The processor (1020) can calculate the similarity between candidate data and one or more features of the representative data. The processor (1020) can calculate the similarity for one or more features.

[0268] In one embodiment, the processor (1020) can select additional training data among candidate data based on the calculated similarity and the similarity range of representative data. The processor (1020) can determine whether to filter the candidate data or use it as additional training data depending on whether the calculated similarity falls within the similarity range of representative data. A specific method of operation is described below with reference to FIGS. 1 to 9.

[0269] In one embodiment, the processor (1020) may update information regarding training data based on additional training data. The processor (1020) may update information regarding training data, such as by adding new representative data or expanding the similarity range of the representative data based on additional training data.

[0270] In one embodiment, the processor (1020) may delete information regarding representative data as it updates information regarding training data. The processor (1020) may delete information regarding representative data based on inclusion relationships between representative data as it updates the similarity range of representative data. As the processor (1020) updates the similarity range of representative data, if the similarity range of the first representative data includes the similarity range of the second representative data, the processor (1020) may delete information regarding the second representative data. For specific operations below, refer to the description of FIG. 8.

[0271] The communication unit (1030) may include one or more modules that enable wireless communication between the electronic device (100) and a network where an external device (not shown) is located. The communication unit (1030) may transmit or receive data or signals to and from an external device (not shown) through a wired or wireless network. A communication unit (1030) according to one embodiment of the present disclosure includes at least one communication module, such as a short-range communication module, a wired communication module, a mobile communication module, a broadcast reception module, etc. Here, the at least one communication module refers to a communication module capable of transmitting and receiving data through a network that follows a communication standard such as a tuner that performs broadcast reception, Bluetooth, WLAN (Wireless LAN) (Wi-Fi), Wibro (Wireless broadband), Wimax (World Interoperability for Microwave Access), CDMA, WCDMA, etc.

[0272] For example, the communication unit (1030) may include a Wi-Fi module, a Bluetooth module, an infrared communication module and a wireless communication module, a LAN module, an Ethernet module, a wired communication module, etc. At this time, each communication module may be implemented in the form of at least one hardware chip. The Wi-Fi module and the Bluetooth module perform communication in the Wi-Fi method and the Bluetooth method, respectively. When using the Wi-Fi module or the Bluetooth module, various connection information such as SSID and session key is first transmitted and received, and after establishing a communication connection using this, various information can be transmitted and received. The wireless communication module may include at least one communication chip that performs communication according to various wireless communication standards such as Zigbee, 3G (3rd Generation), 3GPP (3rd Generation Partnership Project), LTE (Long Term Evolution), LTE-A (LTE Advanced), 4G (4th Generation), 5G (5th Generation), etc. The communication unit (1030) may include a communication unit that performs communication with electronic devices and Bluetooth, and an interface unit that connects to an external device.

[0273] In one embodiment of the present disclosure, the communication unit (1030) may receive candidate data that can be used for training an AI model. The communication unit (1030) may receive candidate data including network data for training an AI model in a wireless communication system.

[0274] A device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, 'non-transitory storage medium' simply means that it is a tangible device and does not contain a signal (e.g., electromagnetic waves), and the term does not distinguish between cases where data is stored semi-permanently and cases where it is stored temporarily. For example, a 'non-transitory storage medium' may include a buffer in which data is stored temporarily.

[0275] According to one embodiment, the method according to the various embodiments disclosed herein may be provided by being included in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed in the form of a device-readable storage medium (e.g., compact disc read-only memory (CD-ROM)), or distributed online (e.g., download or upload) through an application store or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., downloadable app) may be temporarily stored or temporarily created on a device-readable storage medium, such as the memory of a manufacturer's server, an application store's server, or a relay server.

[0276] According to one embodiment of the present disclosure, a method for collecting training data for training an AI model is provided. The method may acquire a plurality of candidate data that can be used for training an AI model. The method may extract one or more features for each of the plurality of candidate data. The method may calculate a similarity between one or more features of at least one representative data included in the existing training data used for training the AI ​​model and one or more features extracted for each of the plurality of candidate data. Based on the calculated similarity and the similarity range of at least one representative data, the method may acquire additional training data among the plurality of candidate data. The method may train an AI model using the additional training data. Based on the additional training data, the method may update information regarding the training data.

[0277] In one embodiment, the similarity range may include one or more characteristic-specific similarity range values.

[0278] In one embodiment, information regarding training data may include information regarding at least one representative data and information regarding a similarity range of at least one representative data.

[0279] In one embodiment, the method can filter the first candidate data based on the calculated similarity when all similarity values ​​between one or more characteristics of the first candidate data included in the plurality of candidate data and one or more characteristics of at least one representative data are included within the similarity range.

[0280] In one embodiment, the method may add the second candidate data as new representative data if, based on the calculated similarity, all similarity values ​​between one or more characteristics of the second candidate data included in the plurality of candidate data and one or more characteristics of at least one representative data are not included within the similarity range.

[0281] In one embodiment, the method can update the similarity range of at least one representative data based on the third candidate data if, based on the calculated similarity, some of the similarity values ​​between one or more characteristics of the third candidate data included in the plurality of candidate data and one or more characteristics of at least one representative data are not included within the similarity range.

[0282] In one embodiment, at least one representative data may include a first representative data and a second representative data. The method may delete information regarding the first representative data if, based on an information update regarding training data, the similarity range of the first representative data and the first representative data is included within the similarity range of the second representative data.

[0283] According to one embodiment of the present disclosure, an electronic device for collecting training data for learning an AI model may include a memory for storing a plurality of instructions and at least one processor for executing a plurality of instructions stored in the memory. By executing a plurality of instructions individually or collectively by at least one processor, the electronic device may acquire a plurality of candidate data that can be used for learning an AI model. By executing a plurality of instructions individually or collectively by at least one processor, the electronic device may extract one or more features for each of the plurality of candidate data. By executing a plurality of instructions individually or collectively by at least one processor, the electronic device may calculate a similarity between one or more features of at least one representative data included in existing training data used for learning an AI model and one or more features extracted for each of the plurality of candidate data. By executing a plurality of instructions individually or collectively by at least one processor, the electronic device may acquire additional training data among the plurality of candidate data based on the calculated similarity and the similarity range of at least one representative data. An electronic device can train an AI model using additional training data by executing a plurality of instructions individually or collectively by at least one processor. An electronic device can update information regarding training data based on additional training data by executing a plurality of instructions individually or collectively by at least one processor.

Claims

1. Regarding the method of collecting training data for training an AI model, A step of acquiring a plurality of candidate data that can be used for training the above AI model; A step of extracting one or more features for each of the above plurality of candidate data; A step of calculating the similarity between one or more features of at least one representative data included in the existing training data used for training the AI ​​model and one or more features extracted for each of the plurality of candidate data; A step of obtaining additional training data among the plurality of candidate data based on the similarity calculated above and the similarity range of the at least one representative data; A step of training the AI ​​model using the additional training data mentioned above; and A method comprising the step of updating information regarding the training data based on the additional training data.

2. In Paragraph 1, A method in which the above similarity range includes similarity range values ​​for one or more of the above characteristics.

3. In either Paragraph 1 or Paragraph 2, A method comprising information regarding the training data above, including information regarding at least one representative data and information regarding the similarity range of the at least one representative data.

4. In any one of paragraphs 1 to 3, The step of acquiring the above additional training data is, A method comprising: a step of filtering the first candidate data when, based on the similarity calculated above, all similarity values ​​between one or more characteristics of the first candidate data included in the plurality of candidate data and one or more characteristics of the at least one representative data are included within the similarity range.

5. In any one of paragraphs 1 through 4, The step of updating information regarding the above training data is, A method comprising the step of adding the second candidate data as new representative data when, based on the similarity calculated above, all similarity values ​​between one or more characteristics of the second candidate data included in the plurality of candidate data and one or more characteristics of the at least one representative data are not included within the similarity range.

6. In any one of paragraphs 1 through 5, The step of updating information regarding the above training data is, A method comprising: a step of updating the similarity range of the at least one representative data based on the third candidate data when, based on the similarity calculated above, some of the similarity values ​​between one or more characteristics of the third candidate data included in the plurality of candidate data and one or more characteristics of the at least one representative data are not included within the similarity range.

7. In any one of paragraphs 1 through 6, The above at least one representative data includes a first representative data and a second representative data, and The above method is, A method further comprising the step of deleting information regarding the first representative data when, based on an update of information regarding the training data, the similarity range of the first representative data and the first representative data is included within the similarity range of the second representative data.

8. In an electronic device (100) for collecting training data for learning an AI model, Memory (1010) for storing multiple instructions; and It includes at least one processor (1020) that executes the plurality of instructions stored in the memory, By executing the above plurality of instructions individually or collectively by the at least one processor (1020), the electronic device (100) Acquire multiple candidate data that can be used to train the above AI model, and One or more features are extracted for each of the above multiple candidate data, and Calculate the similarity between at least one feature of at least one representative data included in the existing training data used for training the above AI model and at least one feature extracted for each of the plurality of candidate data, and Based on the similarity calculated above and the similarity range of at least one representative data, additional training data is obtained among the plurality of candidate data, and The AI ​​model is trained using the additional training data mentioned above, and An electronic device (100) that updates information regarding the training data based on the additional training data above.

9. In Paragraph 8, The above similar range includes similar range values ​​for one or more of the above characteristics, electronic device (100).

10. In either Paragraph 8 or Paragraph 9, The electronic device (100) includes information regarding the training data above, information regarding at least one representative data and information regarding the similarity range of the at least one representative data.

11. In any one of paragraphs 8 through 10, By executing the above plurality of instructions individually or collectively by the at least one processor (1020), the electronic device (100) An electronic device (100) that filters the first candidate data based on the similarity calculated above, when all similarity values ​​between one or more characteristics of the first candidate data included in the plurality of candidate data and one or more characteristics of the at least one representative data are included within the similarity range.

12. In any one of paragraphs 8 through 11, By executing the above plurality of instructions individually or collectively by the at least one processor (1020), the electronic device (100) An electronic device (100) that, based on the similarity calculated above, adds the second candidate data as a new representative data when all similarity values ​​between one or more characteristics of the second candidate data included in the plurality of candidate data and one or more characteristics of the at least one representative data are not included within the similarity range.

13. In any one of paragraphs 8 through 12, By executing the above plurality of instructions individually or collectively by the at least one processor (1020), the electronic device (100) An electronic device (100) that updates the similarity range of the at least one representative data based on the third candidate data when, based on the similarity calculated above, some of the similarity values ​​between one or more characteristics of the third candidate data included in the plurality of candidate data and one or more characteristics of the at least one representative data are not included within the similarity range.

14. In any one of paragraphs 8 through 13, The above at least one representative data includes a first representative data and a second representative data, and By executing the above plurality of instructions individually or collectively by the at least one processor (1020), the electronic device (100) An electronic device (100) that deletes information regarding the first representative data when the similarity range of the first representative data and the similarity range of the first representative data are included within the similarity range of the second representative data, based on an update of information regarding the training data above.

15. A computer-readable recording medium storing a program for executing the method of any one of paragraphs 1 through 7 on a computer.

Images