Data lightening method and system

Abstract

A data lightening method is disclosed. The data lightening method according to one embodiment of the present disclosure may comprise the steps of: embedding each piece of data included in a dataset; determining a support vector of a first class and a support vector of a second class on the basis of the distance between each of the pieces of embedded data; determining a decision boundary on the basis of the support vector of the first class and the support vector of the second class; and determining a first easy sample (ES) farthest from the decision boundary from among the embedded data included in the first class and a second ES farthest from the decision boundary from among the embedded data included in the second class.

Description

Data lightweighting method and system

[0001] A data lightweighting method and a system to which the method is applied are disclosed. More specifically, the present disclosure relates to a method for efficiently forming boundaries of data classes and a system to which the method is applied.

[0002] Recently, computational functions utilizing Artificial Intelligence (AI) have become popularized enough for end users to access. Conventional AI computation involved a user terminal transmitting data related to the task to be computed to an AI model server, and the user terminal receiving the computation results from the server. However, this method caused inconvenience to users because the data communication process between the server and the user terminal took an excessive amount of time.

[0003] Figure 1 illustrates an exemplary data classification method of a Support Vector Machine (SVM) classification model that is commonly used in the past. Referring to Figure 1, although it is possible to efficiently distinguish between data classified into a first class and data classified into a second class using an SVM model, the method based on the conventional SVM model assumes the use of all data classified based on the decision boundary, which has the problem that the amount of data is massive.

[0004] To address the aforementioned problems, engineers in the relevant technical field are seeking methods to embed AI models directly into user terminals; however, since the amount of data required for training AI models is not at a level that personal user terminals can accommodate, a technical solution is currently required.

[0005] Therefore, there has been a demand for a method to drastically reduce the data volume required for training while preserving the computational and inference performance of AI models, but such a method has not been provided due to technical difficulties.

[0006] The technical problem to be achieved through some embodiments of the present disclosure is to provide a method for extracting only data with high discriminative power from data required for training a machine learning model.

[0007] Another technical problem to be achieved through some embodiments of the present disclosure is to provide a method for extracting boundary values ​​for classifying classes from data required for training a machine learning model.

[0008] Another technical objective to be achieved through some embodiments of the present disclosure is to provide a method for calculating the distance between data vectors in a vector space in which data necessary for training a machine learning model is embedded.

[0009] Another technical problem to be achieved through some embodiments of the present disclosure is to provide a method for calculating the distance between data nodes on a knowledge graph in which data necessary for training a machine learning model is graphed.

[0010] Another technical problem to be achieved through some embodiments of the present disclosure is to provide a method for distinguishing nodes corresponding to class boundary values ​​on a knowledge graph in which data necessary for training a machine learning model is graphed.

[0011] The technical problems of the present disclosure are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below.

[0012] A data lightweighting method according to one embodiment of the present disclosure for solving the above technical problem may include the steps of: embedding each of the data included in a dataset; determining a support vector of a first class and a support vector of a second class based on the distance between each of the embedded data; determining a decision boundary based on the support vector of the first class and the support vector of the second class; and determining a first ES (Easy Sample) that is furthest from the decision boundary among the embedded data included in the first class and a second ES that is furthest from the decision boundary among the embedded data included in the second class.

[0013] In some embodiments, the method described above may further include the step of pruning data included in the dataset such that the dataset includes only the support vector of the first class, the support vector of the second class, the first ES, and the second ES.

[0014] In some embodiments, the above method may further include the step of training an artificial neural network based on a dataset obtained as a result of the pruning.

[0015] In some embodiments, the step of determining the first ES and the second ES may include the step of calculating the distance between each of the embedded data included in the first class and the decision boundary. The distance between each of the embedded data included in the first class and the decision boundary may be the value obtained by dividing the absolute value of the difference between each of the embedded data included in the first class and the function value of the decision boundary by the slope value of the decision boundary.

[0016] In some embodiments, the step of determining the first ES and the second ES may include the step of determining a plurality of ES (Easy Samples) of the first class and a plurality of ES of the second class whose distance from the determination boundary is greater than or equal to a reference value.

[0017] In some embodiments, the method described above may further include the step of pruning data included in the dataset such that the dataset includes only the support vector of the first class, the support vector of the second class, a plurality of ESs of the first class, and a plurality of ESs of the second class.

[0018] A data lightweighting method according to another embodiment of the present disclosure for solving the above-mentioned technical problem may include the steps of: generating a knowledge graph based on a dataset; calculating the similarity between each of a plurality of nodes included in the knowledge graph and a first node having a label included in the knowledge graph; classifying nodes with a positive similarity to the first node into a first class and nodes with a negative similarity to the first node into a second class; determining a node among the nodes included in the first class that has the smallest absolute similarity value to the first node as a first reference node; determining a node among the nodes included in the second class that has the smallest absolute similarity value to the first node as a second reference node; determining a node among the nodes included in the first class that has the smallest similarity value to the determined first reference node as a first ES; and determining a node among the nodes included in the second class that has the smallest similarity value to the determined second reference node as a second ES.

[0019] In some embodiments, the method described above may further include the step of pruning the knowledge graph such that the knowledge graph includes only the first reference node, the second reference node, the first ES, and the second ES.

[0020] In some embodiments, the above method may further include the step of training an artificial neural network based on a knowledge graph obtained as a result of the pruning.

[0021] In some embodiments, the step of determining the first ES may include the step of determining a plurality of first class ESs in which the similarity value with the first reference node is less than or equal to a reference value, and the step of determining the second ES may include the step of determining a plurality of second class ESs in which the similarity value with the second reference node is less than or equal to a reference value.

[0022] In some embodiments, the method described above may further include the step of pruning the knowledge graph such that the knowledge graph includes only the first reference node, the second reference node, the plurality of first class ESs, and the plurality of second class ESs.

[0023] A data lightweighting system according to another embodiment of the present disclosure for solving the above-mentioned technical problem may include one or more processors and a memory that stores a computer program executed by said one or more processors. The computer program may enable said processors to perform steps of: embedding each of the data included in a dataset; determining a support vector of a first class and a support vector of a second class based on the distance between each of the embedded data; determining a decision boundary based on the support vector of the first class and the support vector of the second class; and determining a first ES (Easy Sample) that is furthest from the decision boundary among the embedded data included in the first class and a second ES that is furthest from the decision boundary among the embedded data included in the second class.

[0024] In some embodiments, the computer program may further enable the processor to perform the step of pruning data included in the dataset such that the dataset includes only the support vector of the first class, the support vector of the second class, the first ES, and the second ES.

[0025] In some embodiments, the computer program may further enable the processor to perform the step of training an artificial neural network based on the dataset obtained as a result of the pruning.

[0026] In some embodiments, the step of determining the first ES and the second ES may include the step of determining a plurality of ES (Easy Samples) of the first class and a plurality of ES of the second class whose distance from the determination boundary is greater than or equal to a reference value.

[0027] In some embodiments, the computer program may further enable the processor to perform the step of pruning data included in the dataset such that the dataset includes only the support vector of the first class, the support vector of the second class, a plurality of ESs of the first class, and a plurality of ESs of the second class.

[0028] A data lightweighting system according to another embodiment of the present disclosure for solving the above-mentioned technical problem may include one or more processors and a memory for storing a computer program executed by the one or more processors. The above computer program may enable the processor to perform the steps of: generating a knowledge graph based on a dataset; calculating the similarity between each of a plurality of nodes included in the knowledge graph and a first node having a label included in the knowledge graph; classifying nodes with a positive similarity to the first node into a first class and nodes with a negative similarity to the first node into a second class; determining the node with the smallest absolute similarity value to the first node among the nodes included in the first class as a first reference node; determining the node with the smallest absolute similarity value to the first node among the nodes included in the second class as a second reference node; determining the node with the smallest similarity value to the determined first reference node among the nodes included in the first class as a first ES; and determining the node with the smallest similarity value to the determined second reference node among the nodes included in the second class as a second ES.

[0029] In some embodiments, the computer program may further enable the processor to perform the step of pruning the knowledge graph such that the knowledge graph includes only the first reference node, the second reference node, the first ES, and the second ES.

[0030] In some embodiments, the computer program may further enable the processor to perform the step of training an artificial neural network based on a knowledge graph obtained as a result of the pruning.

[0031] In some embodiments, the step of determining the first ES may include the step of determining a plurality of first-class ESs whose similarity value with respect to the first reference node is less than or equal to a reference value. The step of determining the second ES may include the step of determining a plurality of second-class ESs whose similarity value with respect to the second reference node is less than or equal to a reference value.

[0032] In some embodiments, the computer program may further perform the step of pruning the knowledge graph such that the knowledge graph includes only the first reference node, the second reference node, the plurality of first class ESs, and the plurality of second class ESs.

[0033] FIG. 1 is a drawing intended to exemplarily explain the problems of the prior art in the relevant technical field.

[0034] FIG. 2 is a drawing illustrating an exemplary environment in which a data lightweighting system according to one embodiment of the present disclosure can be applied.

[0035] FIG. 3 is a flowchart of a data lightweighting method according to another embodiment of the present disclosure.

[0036] FIG. 4 is a diagram illustrating, by way of example, the steps of determining a support vector and determining a determination boundary that can be performed in some embodiments of the present disclosure.

[0037] FIG. 5 is a drawing for exemplarily illustrating the step of determining a first ES and a second ES that may be performed in some embodiments of the present disclosure.

[0038] FIG. 6 is a drawing for exemplarily illustrating the step of determining a plurality of ESs that can be performed in some embodiments of the present disclosure.

[0039] FIG. 7 is a flowchart of a data lightweighting method according to another embodiment of the present disclosure.

[0040] FIG. 8 is a diagram illustrating, by way of example, the step of generating a knowledge graph based on a dataset that can be performed in some embodiments of the present disclosure.

[0041] FIG. 9 is a diagram illustrating, by way of example, a step of calculating the similarity between each of a plurality of nodes of a knowledge graph and a first node, which can be performed in some embodiments of the present disclosure.

[0042] FIG. 10 is a drawing for exemplarily illustrating the step of determining a first ES and a second ES that may be performed in some embodiments of the present disclosure.

[0043] FIG. 11 is a hardware configuration diagram of a computing system according to another embodiment of the present disclosure.

[0044] Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the attached drawings. The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the attached drawings. However, the technical concept of the present invention is not limited to the following embodiments but can be implemented in various different forms. The following embodiments are provided merely to complete the technical concept of the present invention and to fully inform those skilled in the art of the scope of the present invention, and the technical concept of the present invention is defined only by the scope of the claims.

[0045] In describing the present disclosure, if it is determined that a detailed description of related known configurations or functions could obscure the essence of the invention, such detailed description is omitted.

[0046] Unless otherwise defined, terms used in the following embodiments (including technical and scientific terms) may be used in a meaning commonly understood by those skilled in the art to which this disclosure pertains, but this may vary depending on the intent of those skilled in the art, case law, the emergence of new technology, etc. The terms used in this disclosure are for describing the embodiments and are not intended to limit the scope of this disclosure.

[0047] In the following embodiments, singular expressions include plural concepts unless the context clearly specifies them as singular. Additionally, plural expressions include singular concepts unless the context clearly specifies them as plural.

[0048] In addition, terms such as first, second, A, B, (a), (b), etc. used in the following embodiments are used merely to distinguish one component from another, and the essence, order, or sequence of the said component is not limited by such terms.

[0049] Hereinafter, some embodiments of the present disclosure will be described with reference to the drawings.

[0050] FIG. 2 is a drawing illustrating an exemplary environment in which a data lightweighting system (200) according to one embodiment of the present disclosure can be applied.

[0051] Each component illustrated in FIG. 2 may refer to software or hardware, such as a Field Programmable Gate Array (FPGA) or an Application-Specific Integrated Circuit (ASIC). However, the components are not limited to software or hardware and may be configured to reside in an addressable storage medium or configured to execute one or more processors. Functions provided within the components may be implemented by more subdivided components, or multiple components may be combined to form a single component that performs a specific function.

[0052] The target device (300) shown in FIG. 2 may be a notebook, desktop, laptop, smartphone, and tablet, but is not limited thereto and may include all types of devices equipped with computing functions.

[0053] In some embodiments, the data lightweighting system (200) illustrated in FIG. 2 can communicate with other components through a network. The network can be implemented as any type of wired or wireless network, such as a Local Area Network (LAN), a Wide Area Network (WAN), a mobile radio communication network, or Wibro (Wireless Broadband Internet).

[0054] The data lightweighting system (200) and the database (100) may be understood to operate according to a server-client model. However, in some embodiments, the system may be configured in a client-standalone manner without the need for a server. In this case, the operation performed by the data lightweighting system (200) may be understood to be performed in the database (100).

[0055] Hereinafter, with reference to FIG. 2, the operations that each component illustrated in FIG. 2 can perform will be described.

[0056] A data lightweighting system (200) according to one embodiment of the present disclosure may refer to a computing system that receives a specific dataset containing a plurality of data, selects some of the data included in the received dataset as data to be removed, and performs an operation to reduce the size of the dataset by removing the data to be removed from the dataset. In some embodiments of the present disclosure, the data lightweighting system (200) may apply the dataset with reduced size to an artificial neural network to perform learning on the artificial neural network.

[0057] A database (100) according to one embodiment of the present disclosure may refer to a storage system that stores data to be lightweighted by a data lightweighting system (200). In some embodiments of the present disclosure, the database (100) may be implemented in the form of, for example, a Direct Attached Storage (DAS), Network Attached Storage (NAS), or Storage Area Network (SAN) with the data lightweighting system (200) as the host, but is not limited to any one of these.

[0058] A data lightweighting system (200) according to another embodiment of the present disclosure can receive a specific dataset from a database (100).

[0059] A data lightweighting system (200) according to another embodiment of the present disclosure can embed each of the data included in the dataset. As a result of the above operation, each of the data included in the dataset can be converted into a vector.

[0060] A data lightweighting system (200) according to another embodiment of the present disclosure can determine a first class support vector and a second class support vector based on the distance between each of the embedded data. Support vectors according to some embodiments of the present disclosure can be obviously understood by a person skilled in the art who has knowledge of conventional Support Vector Machine (SVM) classification models.

[0061] In the above-described embodiment, the previously defined classes were exemplified as having a total of two, but in some other embodiments of the present disclosure, the previously defined classes may be three or more.

[0062] A data lightweighting system (200) according to another embodiment of the present disclosure can determine a decision boundary based on the support vectors of the first class and the support vectors of the second class. The decision boundary according to some embodiments of the present disclosure is obvious to a person skilled in the art who has knowledge of conventional support vector machine classification models.

[0063] A data lightweighting system (200) according to another embodiment of the present disclosure can determine the specific data among the embedded data included in the first class that is furthest from the decision boundary as the first ES (Easy Sample).

[0064] In some embodiments of the present disclosure, the ES may mean data that is determined not to be pruned, unlike other data included in the dataset according to specific criteria.

[0065] A data lightweighting system (200) according to another embodiment of the present disclosure can determine the specific data among the embedded data included in the second class that is furthest from the decision boundary as the second ES.

[0066] A data lightweighting system (200) according to another embodiment of the present disclosure can generate a knowledge graph based on the dataset.

[0067] A data lightweighting system (200) according to another embodiment of the present disclosure can calculate the similarity between each of the plurality of nodes included in the knowledge graph and a first node having a label included in the knowledge graph.

[0068] A data lightweighting system (200) according to another embodiment of the present disclosure may classify nodes with a positive similarity to the first node into a first class and nodes with a negative similarity to the first node into a second class.

[0069] A data lightweighting system (200) according to another embodiment of the present disclosure may determine the node that is included in the first class and has the smallest absolute value of similarity with the first node as the first reference node.

[0070] A data lightweighting system (200) according to another embodiment of the present disclosure may be included in the second class, and the node with the smallest absolute value of similarity with the first node may be determined as the second reference node.

[0071] A data lightweighting system (200) according to another embodiment of the present disclosure may determine the node with the smallest similarity value to the determined first reference node among the nodes included in the first class as the first ES.

[0072] A data lightweighting system (200) according to another embodiment of the present disclosure may determine the node with the smallest similarity value to the determined second reference node among the nodes included in the second class as the second ES.

[0073] Up to this point, with reference to FIG. 2, the components included in an exemplary environment where a data lightweighting system (200) can be applied and the operations that said components can perform have been described. The embodiments described above should be understood as exemplary in all respects and not limiting. Furthermore, the configuration and operation of the data lightweighting system (200) according to the present embodiment may be supplemented through several embodiments described later.

[0074] Next, a data lightweighting method according to another embodiment of the present disclosure will be described with reference to FIG. 3. The data lightweighting method according to the present embodiment may be performed by one or more computing systems. Additionally, the data lightweighting method according to the present embodiment may have some operations performed by a first computing device and the remaining operations performed by a second computing device. For example, some operations of the data lightweighting method according to the present embodiment may be performed by an on-premise physical server and the remaining operations may be performed by a cloud compute instance. Additionally, for example, some operations of the data lightweighting method according to the present embodiment may be performed by a data lightweighting system (200) and the remaining operations may be performed by a database (100). Hereinafter, if the entity performing each operation is omitted, it may be understood that the entity performing the operation is the data lightweighting system (200).

[0075] FIG. 3 is a flowchart of a data lightweighting method according to another embodiment of the present disclosure.

[0076] In step S100, the data lightweighting system (200) can embed the dataset into a vector space.

[0077] In some embodiments related to step S100, each of the data included in the dataset can be converted into a vector value by implementing the above-described embodiment.

[0078] In step S200, the data lightweighting system (200) can determine the support vectors for each of the first class and the second class.

[0079] In step S300, the data lightweighting system (200) can determine the determination boundary of the vector space of step S100.

[0080] In some embodiments related to step S300, the data lightweighting system (200) can determine the determination boundary of the vector space of step S100 based on the support vector determined in step S200.

[0081] For example, referring to FIG. 4, the data lightweighting system (200) can obtain an embedded first dataset (41) as a result of embedding the first dataset received from the database (100) in step S100 above into a vector space, and classify each of the data included in the embedded first dataset (41) into a first class or a second class based on a support vector machine algorithm. Additionally, during the process of classifying each of the data included in the embedded first dataset (41) into a first class or a second class, support vectors (43-1, 43-2, 43-3) of the first class of the embedded first dataset (41) and support vectors (44-1, 44-2, 44-3) of the second class of the embedded first dataset (41) can be determined.

[0082] Additionally, the data lightweighting system (200) can determine the first determination boundary (45) of the vector space of the embedded first dataset (41) based on the first class support vectors (43-1, 43-2, 43-3) of the embedded first dataset (41) and the second class support vectors (44-1, 44-2, 44-3) of the embedded first dataset (41).

[0083] In step S400, the data lightweighting system (200) may determine the data corresponding to the vector furthest from the decision boundary among the embedded data included in the first class as the first ES (Easy Sample), and determine the data corresponding to the vector furthest from the decision boundary among the embedded data included in the second class as the second ES.

[0084] For example, referring to FIG. 5, the data lightweighting system (200) can determine the data corresponding to the vector furthest from the first decision boundary (45) among the data corresponding to the vector classified as the first class in the vector space of the embedded first dataset (41) shown in FIG. 5 as the first ES (51).

[0085] For another example, referring to FIG. 5, the data lightweighting system (200) can determine the data corresponding to the vector furthest from the first decision boundary (45) among the data corresponding to the vector classified as the second class in the vector space of the first embedded dataset (41) shown in FIG. 5 as the second ES (52).

[0086] In some embodiments related to step S400, the data lightweighting system (200) can calculate the distance between each of the embedded data and the decision boundary. For example, the data lightweighting system (200) can determine the distance between the first embedded data and the decision boundary by dividing the absolute value of the difference between the function values ​​of the first embedded data and the decision boundary by the slope value of the decision boundary.

[0087] In some other embodiments related to step S400, the data lightweighting system (200) can calculate the distance between each of the embedded data and the decision boundary based on Equation 1.

[0088]

[0089] In mathematical formula 1, is data It can be a label (class), can be a Lagrange multiplier that is always a positive value, and The function is a specific dataset It can mean a mapping function for finding the decision boundary of the vector space corresponding to it.

[0090] In some other embodiments related to step S400, the data lightweighting system (200) may determine a plurality of data corresponding to a plurality of vectors whose distance from the decision boundary in the first class of embedded data is greater than or equal to a reference value as ES.

[0091] For example, referring to FIG. 6, the data lightweighting system (200) determines a support vector (62) corresponding to a first class and a support vector (63) corresponding to a second class in the vector space (60) of the second dataset, determines a second decision boundary (65) of the vector space (60) of the second dataset based on the support vector (62) corresponding to the first class and the support vector (63) corresponding to the second class, and determines a plurality of data corresponding to vectors of the first class that exist at a distance greater than or equal to a reference value from the second decision boundary (65) as a third ES set (61).

[0092] As another example, the data lightweighting system (200) can determine a plurality of data corresponding to vectors of the second class that are at a distance greater than or equal to the second decision boundary (65) of the vector space (60) of the second dataset as the fourth ES set (64).

[0093] In step S500, the data lightweighting system (200) can prune the remaining data of the dataset, excluding the support vectors of each of the first class and second class determined in step S200, and the first ES and second ES determined in step S400.

[0094] In some embodiments related to step S500, the data lightweighting system (200) can prune the remaining data of the dataset excluding the support vectors of each of the first class and the second class determined in step S200 and the plurality of ES determined in step S400.

[0095] For example, referring to FIGS. 4 and 5, the data lightweighting system (200) can prune one or more data other than the first class support vectors (43-1, 43-2, 43-3), the second class support vectors (44-1, 44-2, 44-3), the first ES (51), and the second ES (52) among the data included in the embedded first dataset (41).

[0096] As a result of the above pruning, the embedded first dataset (41) includes support vectors of the first class (43-1, 43-2, 43-3), support vectors of the second class (44-1, 44-2, 44-3), and the first ES (51) and the second ES (52), and other data can be removed from the first dataset (41), and accordingly, the first dataset (41) can be made lightweight.

[0097] For another example, referring to FIG. 6, the data lightweighting system (200) can prune one or more data other than the support vectors (62) corresponding to the first class, the support vectors (63) corresponding to the second class, the plurality of data included in the third ES set (61), and the plurality of data included in the fourth ES set (64) among the data included in the vector space (60) of the second dataset.

[0098] As a result of the above pruning, the second dataset includes support vectors (62) corresponding to the first class, support vectors (63) corresponding to the second class, multiple data included in the third ES set (61), and multiple data included in the fourth ES set (64), and other data may be removed from the second dataset, and accordingly, the second dataset may be lightweight.

[0099] In some other embodiments related to step S500, the data lightweighting system (200) can train an artificial neural network based on the dataset obtained as a result of the pruning.

[0100] According to some embodiments of the present disclosure, an artificial neural network may be supervised, semi-supervised, or unsupervised based on a dataset (i.e., a training set). For example, a dataset comprising at least one of a plurality of positive prompts, a plurality of negative prompts, and a plurality of ground truth data may be applied to an artificial neural network, and the artificial neural network may be supervised so that an image (or feature data) output from the artificial neural network based on the ground truth data among the positive prompts and negative prompts becomes similar to / identical to the texture (or feature data) included in the ground truth data. When the artificial neural network undergoes repeated learning, the weights of each node included in the artificial neural network may converge to an optimal value.

[0101] Additionally, the data lightweighting system (200) transmits the learned artificial neural network to the target device (300), and the target device (300) may perform a specific task based on the received artificial neural network or update a dataset corresponding to the artificial neural network already existing in the target device (300).

[0102] According to the above embodiment, as illustrated in FIG. 6, the data lightweighting system (200) can define the regions of the first class and the second class using only the support vector (62) corresponding to the first class of the vector space (60) of the second dataset, the support vector (63) corresponding to the second class, the plurality of data included in the third ES set (61), and the plurality of data included in the fourth ES set (64).

[0103] In the case of conventional technology, there was a problem in that all data included in the dataset had to be used to determine the region or range of each class. According to the present embodiment, the data lightweighting system (200) may achieve the effect of building a machine learning model with performance similar to conventional technology, which defines the region of a class using all data included in the dataset, with less data.

[0104] Up to now, a data lightweighting method according to another embodiment of the present disclosure has been described with reference to FIGS. 3 through 6. The embodiments described above should be understood as illustrative in all respects and not restrictive.

[0105] Hereinafter, a data lightweighting method according to another embodiment of the present disclosure will be described with reference to FIGS. 7 to 10. Unless otherwise noted, the steps described in some flowcharts below may be understood to be performed by the data lightweighting system (200) described with reference to FIG. 2. Furthermore, it is obvious that the technical concept understood in the embodiment described above with reference to FIG. 2 can be applied to the data lightweighting method according to the present embodiment.

[0106] In step S700 illustrated in FIG. 7, the data lightweighting system (200) can generate a knowledge graph based on a dataset received from the database (100).

[0107] For example, referring to FIG. 8, the data lightweighting system (200) can generate a knowledge graph (81) of a third dataset based on a third dataset received from a database (100). Here, some nodes of the knowledge graph, such as the first node (82) shown in FIG. 8, may be nodes that already have labels.

[0108] In step S800, the data lightweighting system (200) can calculate the similarity between each node of the knowledge graph generated in step S700 and the first node having a label. Here, the operation of calculating the similarity may refer to the calculation of similarity between nodes within the graph using a conventional random walk, but is not limited to any single conventional technology as long as it is a method capable of calculating similarity or distance between nodes within the knowledge graph.

[0109] For example, referring to FIG. 9, the data lightweighting system (200) can calculate the similarity between the first node (82) of the knowledge graph (81) of the third dataset and each of the other nodes included in the knowledge graph (81) of the third dataset. Additionally, based on the calculated similarity, each node of the knowledge graph (81) of the third dataset can be labeled.

[0110] In step S900, the data lightweighting system (200) can classify the nodes included in the knowledge graph into a specific class based on the similarity between each node of the knowledge graph and the first node.

[0111] For example, referring to FIG. 9, the data lightweighting system (200) can classify the second node (83) into the same first class as the first node (82) when the similarity between the second node (83) and the first node (82) is positive.

[0112] In another example, the data lightweighting system (200) can classify the third node (84) into a second class different from the first node (82) when the similarity between the third node (84) and the first node (82) is negative.

[0113] In step S1000, the data lightweighting system (200) can determine a first reference node and a second reference node among the nodes included in the knowledge graph.

[0114] In some embodiments related to step S1000, the data lightweighting system (200) may determine the node with the smallest absolute value of similarity with the first node among the nodes included in the first class of the knowledge graph as the first reference node.

[0115] In some other embodiments related to step S1000, the data lightweighting system (200) may determine the node with the smallest similarity value with the first node among the nodes with a positive similarity value with the first node among the plurality of nodes included in the knowledge graph as the first reference node.

[0116] In some other embodiments related to step S1000, the data lightweighting system (200) may determine the node with the smallest absolute value of similarity with the first node among the nodes included in the second class of the knowledge graph as the second reference node.

[0117] In some other embodiments related to step S1000, the first reference node may refer to a node that serves as a reference for a similarity operation to determine ES among the nodes included in the first class of the knowledge graph, but this will be described later.

[0118] For example, referring to FIG. 9, the data lightweighting system (200) can determine the second node (83), which is the node of the first class with the smallest similarity value with the first node (82) included in the knowledge graph (81) of the third dataset, as the first reference node.

[0119] For another example, referring to FIG. 9, the data lightweighting system (200) can determine the third node (84), which is the node with the largest similarity value with the first node (82) among the nodes included in the second class of the knowledge graph (81) of the third dataset, as the second reference node.

[0120] In step S1100, the data lightweighting system (200) can determine the ES of the knowledge graph based on the first reference node and the second reference node determined in step S1000.

[0121] In some embodiments related to step S1100, the data lightweighting system (200) may determine the node with the smallest similarity value to the determined first reference node among a plurality of nodes included in the first class of the knowledge graph as the first ES.

[0122] For example, referring to FIG. 10, when the first reference node of the knowledge graph (101) of the fourth dataset is the fourth node (108), the data lightweighting system (200) can determine the eighth node (109), which has the smallest similarity value with the fourth node (108) and the greatest distance between nodes, as the first ES.

[0123] In some other embodiments related to step S1100, the data lightweighting system (200) may determine the node with the smallest similarity value to the determined second reference node among a plurality of nodes included in the second class of the knowledge graph as the second ES.

[0124] For example, referring to FIG. 10, when the second reference node of the knowledge graph (101) of the fourth dataset is the sixth node (103), the data lightweighting system (200) can determine the seventh node (102), which has the smallest similarity value with the sixth node (103), as the second ES. Here, the fact that the seventh node (102) has the smallest similarity value with the sixth node (103) among the multiple nodes included in the second class of the knowledge graph (101) of the fourth dataset may mean that the seventh node (102) is the node that is furthest from the sixth node (103) on the knowledge graph (101) of the fourth dataset.

[0125] In some other embodiments related to step S1100, the first reference node and the second reference node may exist in multiple quantities, and the ES of the knowledge graph may also exist in multiple quantities.

[0126] In some other embodiments related to step S1100, the data lightweighting system (200) may determine all nodes of a plurality of first class, whose similarity value with the first reference node is less than or equal to a reference value, as ES.

[0127] In some other embodiments related to step S1100, the data lightweighting system (200) may determine all nodes furthest from each of the first reference nodes as ES when there are multiple first reference nodes.

[0128] For example, referring to FIG. 10, when the first reference node of the knowledge graph (101) of the fourth dataset is the fourth node (108), the data lightweighting system (200) can determine both the eighth node (109) and the ninth node (112), which are nodes of the first class whose similarity to the fourth node (108) is less than or equal to the reference value, as the ES of the knowledge graph (101) of the fourth dataset.

[0129] For another example, referring to FIG. 10, when the first reference node of the knowledge graph (101) of the fourth dataset is the fourth node (108), the tenth node (107), and the eleventh node (106), the data lightweighting system (200) can determine the eighth node (109), which has the smallest similarity to the fourth node (108) and the tenth node (107), and the ninth node (112), which has the smallest similarity to the eleventh node (106), as the ES of the knowledge graph (101) of the fourth dataset.

[0130] For another example, referring to FIG. 10, when the second reference node of the knowledge graph (101) of the fourth dataset is the sixth node (103), the data lightweighting system (200) can determine both the seventh node (102) and the twelfth node (111), which are nodes of the second class whose similarity to the sixth node (103) is less than or equal to the reference value, as the ES of the knowledge graph (101) of the fourth dataset.

[0131] In another example, referring to FIG. 10, when the second reference node of the knowledge graph (101) of the fourth dataset is the sixth node (103), the thirteenth node (104), and the fourteenth node (105), the data lightweighting system (200) can determine the seventh node (102), which has the smallest similarity to the sixth node (103) and the thirteenth node (104), as the ES of the knowledge graph (101) of the fourth dataset, and the twelfth node (111), which has the smallest similarity to the fourteenth node (105), as the ES of the knowledge graph (101) of the fourth dataset.

[0132] In step S1200, the data lightweighting system (200) can prune one or more nodes other than the reference node determined in step S1000 and the ES determined in step S1100 from the knowledge graph.

[0133] As a result of the above pruning, the knowledge graph includes the first reference node, the second reference node, and the ES determined in step S1100, and other data can be removed from the knowledge graph, and accordingly, the dataset corresponding to the knowledge graph can be lightweighted.

[0134] Up to this point, a data lightweighting method according to another embodiment of the present disclosure has been described with reference to FIGS. 7 through 10. The embodiments described above should be understood as illustrative in all respects and not limiting.

[0135] FIG. 11 is a hardware configuration diagram of a computing system (1000) according to some embodiments of the present disclosure. The computing system (1000) of FIG. 11 may refer to a data lightweighting system (200) described, for example, with reference to FIG. 2. The computing system (1000) may include one or more processors (1100), a system bus (1600), a communication interface (1200), a memory (1400) for loading a computer program (1500) executed by the processor (1100), and a storage (1300) for storing the computer program (1500).

[0136] The processor (1100) controls the overall operation of each component of the computing system (1000). The processor (1100) may perform operations on at least one application or program for executing methods / operations according to various embodiments of the present disclosure. Memory (1400) stores various data, instructions and / or information. Memory (1400) may load one or more computer programs (1500) from storage (1300) to execute methods / operations according to various embodiments of the present disclosure. A bus (1600) provides communication functions between components of the computing system (1000). A communication interface (1200) supports internet communication of the computing system (1000). Storage (1300) may non-temporarily store one or more computer programs (1500). A computer program (1500) may include one or more instructions in which methods / operations according to various embodiments of the present disclosure are implemented. When a computer program (1500) is loaded into memory (1400), the processor (1100) can perform methods / operations according to various embodiments of the present disclosure by executing one or more of the instructions.

[0137] In some embodiments, the computing system (1000) described with reference to FIG. 11 may be configured using one or more physical servers included in a server farm based on cloud technology such as a virtual machine. In this case, at least some of the components shown in FIG. 11, such as the processor (1100), memory (1400), and storage (1300), may be virtual hardware, and the communication interface (1200) may also be configured as a virtualized networking element such as a virtual switch.

[0138] A computer program (1500) according to some embodiments of the present disclosure may include instructions for performing steps of: embedding each of the data included in a dataset of a computing system (1000); determining a support vector of a first class and a support vector of a second class based on the distance between each of the embedded data; determining a decision boundary based on the support vector of the first class and the support vector of the second class; and determining a first ES (Easy Sample) that is furthest from the decision boundary among the embedded data included in the first class and a second ES that is furthest from the decision boundary among the embedded data included in the second class.

[0139] Various embodiments of the present disclosure and effects according to those embodiments have been described with reference to FIGS. 1 to 11. The effects according to the technical concept of the present disclosure are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below.

[0140] The technical concept of the present disclosure described above may be implemented as computer-readable code on a computer-readable medium. The computer program recorded on the computer-readable recording medium may be transmitted to another computing device via a network such as the Internet and installed on the other computing device, thereby being used on the other computing device.

[0141] Although operations are depicted in a specific order in the drawings, it should not be understood that the operations must necessarily be executed in the specific order depicted or in a sequential order, or that all depicted operations must be executed to obtain the desired result. In certain situations, multitasking and parallel processing may be advantageous. Although embodiments of the present disclosure have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention may be practiced in other specific forms without altering the technical concept or essential features thereof. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. The scope of protection of the present invention shall be interpreted by the claims below, and all technical concepts within the equivalent scope shall be interpreted as being included within the scope of rights of the technical concept defined by the present disclosure.

Claims

1. In a method performed by a computing system, A step of embedding each piece of data included in the dataset; A step of determining a support vector of the first class and a support vector of the second class based on the distance between each of the above-mentioned embedded data; A step of determining a decision boundary based on the support vectors of the first class and the second class; and A step comprising determining a first ES (Easy Sample) that is furthest from the decision boundary among the embedded data included in the first class, and a second ES that is furthest from the decision boundary among the embedded data included in the second class. Data lightweighting method.

2. In Paragraph 1, The method further comprises the step of pruning data included in the dataset such that the dataset includes only the support vector of the first class, the support vector of the second class, the first ES, and the second ES. Data lightweighting method.

3. In Paragraph 2, A method further comprising the step of training an artificial neural network based on a dataset obtained as a result of the above pruning. Data lightweighting method.

4. In Paragraph 1, The step of determining the first ES and the second ES is, The method includes the step of calculating the distance between each of the embedded data included in the first class and the decision boundary. The distance between each of the embedded data included in the first class and the decision boundary is, The value obtained by dividing the absolute value of the difference between each of the embedded data included in the first class and the function value of the decision boundary by the slope value of the decision boundary, Data lightweighting method.

5. In Paragraph 1, The step of determining the first ES and the second ES is, A step comprising determining a plurality of ES (Easy Samples) of the first class and a plurality of ES of the second class, wherein the distance from the decision boundary is greater than or equal to a reference value. Data lightweighting method.

6. In Paragraph 5, The method further comprises the step of pruning data included in the dataset such that the dataset includes only the support vectors of the first class, the support vectors of the second class, a plurality of ESs of the first class, and a plurality of ESs of the second class. Data lightweighting method.

7. In a method performed by a computing system, A step of generating a knowledge graph based on a dataset; A step of calculating the similarity between each of the plurality of nodes included in the knowledge graph and the first node having a label included in the knowledge graph; A step of classifying nodes with a positive similarity to the first node into a first class, and nodes with a negative similarity to the first node into a second class; A step of determining the node with the smallest absolute value of similarity with the first node among the nodes included in the first class as the first reference node; A step of determining the node with the smallest absolute similarity value with the first node among the nodes included in the second class as the second reference node; A step of determining the node with the smallest similarity value to the determined first reference node among the nodes included in the first class as the first ES; and A step comprising determining the node with the smallest similarity value to the determined second reference node among the nodes included in the second class as the second ES. Data lightweighting method.

8. In Paragraph 7, The method further comprises the step of pruning the knowledge graph such that the knowledge graph includes only the first reference node, the second reference node, the first ES, and the second ES. Data lightweighting method.

9. In Paragraph 8, A method further comprising the step of training an artificial neural network based on a knowledge graph obtained as a result of the above pruning. Data lightweighting method.

10. In Paragraph 7, The step of determining the first ES above is, It includes the step of determining a plurality of ESs of the first class whose similarity value with the first reference node is less than or equal to a reference value, The step of determining the second ES above is, A step comprising determining a plurality of ESs of a second class whose similarity value with the second reference node is less than or equal to a reference value, Data lightweighting method.

11. In Paragraph 10, The method further comprises the step of pruning the knowledge graph such that the knowledge graph includes only the first reference node, the second reference node, the plurality of first class ESs, and the plurality of second class ESs. Data lightweighting method.

12. One or more processors; and It includes memory for storing computer programs executed by one or more of the above processors, and The above computer program causes the processor, A step of embedding each piece of data included in the dataset; A step of determining a support vector of the first class and a support vector of the second class based on the distance between each of the above-mentioned embedded data; A step of determining a decision boundary based on the support vectors of the first class and the second class; and A step of determining the first ES (Easy Sample) that is furthest from the decision boundary among the embedded data included in the first class, and the second ES that is furthest from the decision boundary among the embedded data included in the second class. Data lightweighting system.

13. In Paragraph 12, The above computer program causes the processor, Further performing the step of pruning the data included in the dataset such that the dataset includes only the support vector of the first class, the support vector of the second class, the first ES, and the second ES. Data lightweighting system.

14. In Paragraph 13, The above computer program causes the processor, Further performing the step of training an artificial neural network based on the dataset obtained as a result of the above pruning, Data lightweighting system.

15. In Paragraph 12, The step of determining the first ES and the second ES is, A step comprising determining a plurality of ES (Easy Samples) of the first class and a plurality of ES of the second class, wherein the distance from the decision boundary is greater than or equal to a reference value. Data lightweighting system.

16. In Paragraph 15, The above computer program causes the processor, Further performing the step of pruning data included in the dataset such that the dataset includes only the support vector of the first class, the support vector of the second class, a plurality of ESs of the first class, and a plurality of ESs of the second class. Data lightweighting system.

17. One or more processors; It includes memory for storing computer programs executed by one or more of the above processors, and The above computer program causes the processor, A step of generating a knowledge graph based on a dataset; A step of calculating the similarity between each of the plurality of nodes included in the knowledge graph and the first node having a label included in the knowledge graph; A step of classifying nodes with a positive similarity to the first node into a first class, and nodes with a negative similarity to the first node into a second class; A step of determining the node with the smallest absolute value of similarity with the first node among the nodes included in the first class as the first reference node; A step of determining the node with the smallest absolute similarity value with the first node among the nodes included in the second class as the second reference node; A step of determining the node with the smallest similarity value to the determined first reference node among the nodes included in the first class as the first ES; and Among the nodes included in the second class, performing the step of determining the node with the smallest similarity value to the determined second reference node as the second ES. Data lightweighting system.

18. In Paragraph 17, The above computer program causes the processor, Further performing the step of pruning the knowledge graph such that the knowledge graph includes only the first reference node, the second reference node, the first ES, and the second ES, Data lightweighting system.

19. In Paragraph 18, The above computer program causes the processor, Further performing the step of training an artificial neural network based on the knowledge graph obtained as a result of the above pruning, Data lightweighting system.

20. In Paragraph 17, The step of determining the first ES above is, It includes the step of determining a plurality of ESs of the first class whose similarity value with the first reference node is less than or equal to a reference value, The step of determining the second ES above is, A step comprising determining a plurality of ESs of a second class whose similarity value with the second reference node is less than or equal to a reference value, Data lightweighting system.

21. In Paragraph 20, The above computer program causes the processor, Further performing the step of pruning the knowledge graph such that the knowledge graph includes only the first reference node, the second reference node, the plurality of first class ESs, and the plurality of second class ESs. Data lightweighting system.

Images