KR102249818B1 - Method and apparatus of generating training data based on distribution characteristics of training data - Google Patents

Abstract

A method and apparatus for generating training data according to an embodiment divides training data to be oversampled into isolated training data and normal training data based on distribution characteristics of training data, and determines the ratio between the isolated training data and the normal training data. According to this, the number of additional training data to be oversampled is determined from each of the isolated training data and the normal training data, and additional training data corresponding to each of the isolated training data and the normal training data is generated based on the number of each of the training data to be oversampled. do.

Description

Method and apparatus for generating training data based on the distribution characteristics of training data {METHOD AND APPARATUS OF GENERATING TRAINING DATA BASED ON DISTRIBUTION CHARACTERISTICS OF TRAINING DATA}

The following embodiments relate to a method and apparatus for generating training data based on distribution characteristics of the training data.

In order to create a classifier with good performance, training data used for learning artificial intelligence is very important. The performance of the classifier may show a large difference depending on, for example, the quality and number of training data, and a ratio between training data.

In general, the more the number of training data, the better the performance of the classifier, but the ratio between training data is also very important. For example, if the number of training data for one class is large and the number of training data for another class is small, a class with a large number of training data can learn relatively well, but a class with insufficient number of training data is Learning may not work well. When the number of training data or the ratio between training data is not appropriate as described above, the performance of the classifier may be biased.

According to an embodiment, the number of training data for a classifier or a ratio between training data may be appropriately adjusted by oversampling training data based on distribution characteristics of the training data.

According to an embodiment, reliable training data may be generated by appropriately adjusting the number of training data for a classifier or a ratio between training data, and a highly reliable classifier may be provided.

According to one side, a method of generating training data includes dividing training data to be oversampled into isolation training data and normal training data based on distribution characteristics of the training data; Determining the number of additional learning data to be oversampled from each of the isolated learning data and the normal learning data according to a ratio between the isolated learning data and the normal learning data; And generating additional training data corresponding to each of the isolation training data and the normal training data, based on the number of each of the training data to be oversampled.

The discriminating step is based on a ratio between training data of the target class to be oversampled among the training data for different classes and additional training data to be oversampled from the training data of the target class, Extracting closest training data for each training data of the target class; And dividing the learning data of the target class into the isolation learning data and the normal learning data based on the nearest learning data for each of the learning data of the target class.

The extracting of the nearest training data for each of the training data of the target class includes training data of the target class and additional training data to be oversampled based on a distance between each of the training data of the target class. It may include the step of extracting the nearest learning data corresponding to the ratio between.

In the step of dividing the learning data of the target class into the isolation learning data and the normal learning data based on the nearest learning data, all of the nearest learning data are learning data of the target class and heterogeneous learning data. In case, dividing learning data of a target class corresponding to the nearest learning data into the isolated learning data; And when at least one of the closest training data is training data of the same kind as the training data of the target class, dividing training data of the target class corresponding to the closest training data into the normal training data. can do.

The determining of the number of additional learning data includes a first number of additional learning data to be oversampled from the isolated learning data and an addition to be oversampled from the normal learning data according to a ratio between the isolation learning data and the normal learning data. It may include determining a second number of training data.

The generating of the additional learning data includes extracting the additional learning data based on a first number of additional learning data to be oversampled from the isolated learning data and a second number of additional learning data to be oversampled from the normal learning data. The step of doing; And a first distance between the isolated learning data and the additional learning data extracted from the isolated learning data, and a second distance between the normal learning data and the additional learning data extracted from the normal learning data, respectively, the isolated learning data, and And generating additional learning data corresponding to each of the normal learning data.

The extracting of the additional learning data may include extracting a third number of additional learning data from the learning data of the target class based on the first number of additional learning data to be oversampled from the isolated learning data; And extracting a fourth number of additional training data from the training data based on the second number of additional training data to be oversampled from the normal training data.

The generating of the additional learning data is based on each of a first distance between the isolation learning data and the third number of additional learning data and a second distance between the normal learning data and the fourth number of additional learning data. And calculating the number of samplings per unit distance of each of the isolated learning data and the normal learning data; And generating additional learning data corresponding to each of the isolation learning data and the normal learning data, based on the first distance, the second distance, and the number of samplings per unit distance of each of the isolation learning data and the normal learning data. It may include steps.

The calculating of the number of samplings per unit distance of each of the isolation learning data and the normal learning data may include calculating a first distance between the isolation learning data and the third number of additional learning data; And calculating the number of samplings per unit distance of the isolation learning data based on the first distance.

The generating of the additional learning data may include oversampling additional learning data corresponding to the isolated learning data in inverse proportion to the first distance.

The oversampling of the additional learning data corresponding to the isolated learning data is a number less than the number of samples per unit distance of the isolated learning data from the isolated learning data when the first distance is greater than a unit distance of the isolated learning data. Oversampling the additional learning data of; And when the first distance is equal to or close to the unit distance of the isolation learning data, oversampling a number of additional learning data greater than the number of samples per unit distance of the isolation learning data from the isolation learning data. have.

The generating of the additional learning data may include determining whether additional learning data corresponding to the isolated learning data and heterogeneous learning data overlap; And removing additional learning data corresponding to the isolated learning data based on a result of the determination as to whether or not to overlap.

The calculating of the number of samplings per unit distance of each of the isolated learning data and the normal learning data may include calculating a second distance between the normal learning data and the fourth number of additional learning data; And calculating the number of samplings per unit distance of the normal learning data based on the second distance.

The generating of the additional learning data may include oversampling additional learning data corresponding to the normal learning data in proportion to the second distance.

The oversampling of the additional training data corresponding to the normal training data includes a number greater than the number of samples per unit distance of the normal training data from the normal training data when the second distance is greater than the unit distance of the normal training data. Oversampling the additional learning data of; And when the second distance is equal to or close to the unit distance of the normal learning data, oversampling a number of additional training data less than the number of samples per unit distance of the normal learning data from the normal learning data. have.

According to one side, the apparatus for generating training data divides training data to be oversampled into isolated training data and normal training data based on the distribution characteristics of the training data, and determines the ratio between the isolated training data and the normal training data. Accordingly, the number of additional learning data to be oversampled is determined from each of the isolated learning data and the normal learning data, and based on the number of each of the learning data to be oversampled, each of the isolated learning data and the normal learning data is And a processor that generates corresponding additional learning data.

The processor learns the target class based on a ratio between training data of the target class to be oversampled among the training data for different classes and additional training data to be oversampled from the training data of the target class. The closest learning data for each data is extracted, and based on the nearest learning data for each of the training data of the target class, the learning data of the target class can be divided into the isolation learning data and the normal learning data. have.

The processor may extract closest training data corresponding to a ratio between training data of the target class and additional training data to be oversampled based on a distance between each of the training data of the target class.

When all of the nearest learning data are learning data of the target class and learning data of a different type, the processor divides learning data of a target class corresponding to the nearest learning data into the isolation learning data, and When at least one of the contact training data is training data of the same kind as the training data of the target class, training data of a target class corresponding to the closest training data may be classified as the normal training data.

The processor determines a first number of additional learning data to be oversampled from the isolated learning data and a second number of additional learning data to be oversampled from the normal learning data according to a ratio between the isolated learning data and the normal learning data. I can.

The processor extracts the additional learning data based on a first number of additional learning data to be oversampled from the isolated learning data and a second number of additional learning data to be oversampled from the normal learning data, and the isolated learning data And a first distance between the additional learning data extracted from the isolated learning data and a second distance between the normal learning data and the additional learning data extracted from the normal learning data, respectively, the isolation learning data and the normal learning data, respectively. It is possible to generate additional learning data corresponding to.

According to one side, the number of training data for the classifier or a ratio between training data may be appropriately adjusted.

According to one side, it is possible to generate highly reliable training data and provide a highly reliable classifier.

1 is a flow chart showing a method of generating training data according to an embodiment.
2 is a view for explaining isolation learning data and normal learning data according to an embodiment.
3 is a diagram for describing a method of generating additional learning data corresponding to isolation learning data according to an embodiment.
FIG. 4 is a diagram for explaining a method of removing additional learning data according to whether the additional learning data generated in response to the isolated learning data and the learning data are overlapped, according to an exemplary embodiment.
5 is a diagram for describing a method of generating additional learning data corresponding to normal learning data according to an embodiment.
6 is a flowchart illustrating a method of generating learning data according to another embodiment.
Fig. 7 is a block diagram of an apparatus for generating training data according to an exemplary embodiment.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. However, the scope of the patent application is not limited or limited by these embodiments. The same reference numerals shown in each drawing indicate the same members.

Various changes may be made to the embodiments described below. The embodiments described below are not intended to be limited to the embodiments, and should be understood to include all changes, equivalents, and substitutes thereto.

The terms used in the examples are used only to describe specific embodiments, and are not intended to limit the embodiments. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present specification, terms such as "comprise" or "have" are intended to designate the presence of features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, but one or more other features. It is to be understood that the presence or addition of elements or numbers, steps, actions, components, parts, or combinations thereof does not preclude in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiment belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in the present application. Does not.

In addition, in the description with reference to the accompanying drawings, the same reference numerals are assigned to the same components regardless of the reference numerals, and redundant descriptions thereof will be omitted. In describing the embodiments, when it is determined that a detailed description of related known technologies may unnecessarily obscure the subject matter of the embodiments, the detailed description thereof will be omitted.

1 is a flowchart illustrating a method of generating learning data according to an exemplary embodiment. Referring to FIG. 1, an apparatus for generating training data (hereinafter, a “generation device”) according to an embodiment isolates training data to be oversampled based on distribution characteristics of training data. And normal training data (110). The training data may include, for example, training data for different classes. The training data may include, for example, first training data for learning of a first class and second training data for learning of a second class.

The generating device is based on the ratio between the training data of the target class to be oversampled among the training data and the additional training data to be oversampled from the training data of the target class, the nearest learning data for each of the training data of the target class. Can be extracted. Here, the'target class' may be understood as a class corresponding to training data to be oversampled due to, for example, an insufficient number of training data. In addition,'oversampling' is to increase the number of training data, and may be understood as having the same meaning as sampling or additional generation.

The generating device may extract closest learning data corresponding to a ratio between the training data of the target class and the additional training data to be oversampled based on the distance between each of the training data of the target class.

For example, suppose that the second training data among the training data is oversampled. In this case, the class of the second learning data may be the target class. In addition, suppose that the number of second training data is N and the number of additional training data to be oversampled from the N second training data is S.

The generating device is based on the ratio (eg, S/N) between the N second training data and the S additional training data to be oversampled from the N second training data, the training data of the target class ( The nearest learning data for each of the second learning data) may be extracted. For example, if N is 10 and S is 60, since S/N = 6, the generating device may extract 6 nearest training data for each of the second training data. In this case, the nearest learning data are learning data adjacent to any one of the second learning data (for example, a learning data), and may be either the first learning data or the second learning data. The generating device may extract 6 pieces of training data at the closest distance as nearest training data based on the distance between the training data a and the adjacent training data. Here, the'distance' may correspond to an Euclidian distance indicating similarity between training data.

The generating device may classify the learning data of the target class into isolation learning data and normal learning data based on the extracted S/N nearest learning data. For example, when all of the six closest learning data are learning data of a target class (second learning data) and heterogeneous learning data (first learning data), the target corresponding to the nearest learning data Class learning data (a learning data) can be classified as isolated learning data. Here, the training data of the target class corresponding to the nearest training data may correspond to training data adjacent to all of the nearest training data or training data closest to all of the nearest training data. Alternatively, when at least one of the nearest learning data is learning data of the target class (second learning data) and the same kind of learning data (second learning data), the generation device learns the target class corresponding to the nearest learning data. Data can be classified as normal learning data.

For example, since the number of second training data is N, _{if the number of isolated training data is N 1} , the number of normal training data may be _{N 2.} In other words. N = N ₁ + N ₂

The generating device may automatically determine a ratio between the isolated learning data and the normal learning data among additional learning data to be finally oversampled according to a ratio between the isolated learning data and the normal learning data.

The isolation learning data and the normal learning data will be described in detail with reference to FIG. 2.

The generating device determines the number of additional learning data to be oversampled from each of the isolated learning data and the normal learning data according to a ratio between the isolated learning data and the normal learning data (120). The generating device may determine a first number of additional training data to be oversampled from the isolated training data and a second number of additional training data to be oversampled from the normal training data according to a ratio between the isolated training data and the normal training data.

For example, suppose that the number of isolated training data among training data to be oversampled (second training data) is N ₁ and the number of normal training data is N ₂ . The generating device is based on the number of isolated learning data (N ₁ ), the number of normal training data (N ₂ ), and the number of additional training data to be oversampled from each of the isolated learning data and the normal learning data (S). A first number (S ₁ ) of additional training data to be oversampled from the data and a second number (S ₂ ) of additional training data to be oversampled from the normal training data may be determined.

For example, the generating device uses the ratio (N ₁ /N) of the isolated learning data to set the first number (S ₁ ) of _{additional learning data to be oversampled from the isolated learning data S 1} = S*(N ₁ /N) ). In addition, the generation device uses the ratio of the normal learning data (N _{2 /N) to determine the second number (S 2} ) of additional training data to be oversampled from the normal training data _{, S 2} = S*(N ₂ /N) and You can get it together.

The generating device generates additional learning data corresponding to each of the isolated learning data and the normal learning data based on the number of each of the learning data to be oversampled (130).

The generating device may extract additional training data based on the first number of additional training data to be oversampled from the isolated training data and the second number of additional training data to be oversampled from the normal training data.

_{The generating device is, for example, based on the first number (S 1} ) of additional training data to be oversampled from the isolated training data, a third number (eg, S ₁ /N _{1) from the N second training data.} Dog) of the nearest additional learning data can be extracted. _{In addition, the generation device is based on the second number (S 2} ) of the additional training data to be oversampled from the normal training data, based on the fourth number (eg, from all training data (first training data + second training data)). For example, S ₂ /N ₂ ) of the nearest additional learning data can be extracted.

In addition, the generation device is based on a first distance between the isolation learning data and the additional learning data extracted from the isolation learning data, and the second distance between the normal learning data and the additional learning data extracted from the normal learning data, respectively, based on the isolation learning data and the normal. Additional training data corresponding to each of the training data may be generated.

The generating device is based on the first distance between the isolation learning data and the third number of additional learning data and the second distance between the normal learning data and the fourth number of additional learning data, respectively, the isolation learning data and the normal learning data. The number of samples per unit distance can be calculated. The generating device may generate additional learning data corresponding to each of the isolation learning data and the normal learning data, based on the first distance, the second distance, and the number of samplings per unit distance of each of the isolated learning data and the normal learning data.

A method of generating additional learning data corresponding to the isolation learning data by the generation device will be described in detail with reference to FIGS. 3 to 4. In addition, a method of generating the additional learning data corresponding to the normal learning data by the generation device will be described in detail with reference to FIG. 6.

2 is a diagram illustrating isolation learning data and normal learning data according to an embodiment. Referring to FIG. 2, distributions of first training data indicated by a circle and second training data indicated by a triangle are shown.

For example, it is assumed that the training data of the target class to be oversampled are second training data.

If all of the nearest learning data included in the area 210, such as the learning data a, among the second learning data displayed in a triangular form, are the learning data of the target class and the learning data of the target class (first learning data), the generating device May classify learning data (a learning data) of a target class corresponding to the nearest learning data as isolated learning data.

In addition, at least one of the nearest learning data, such as b learning data included in the area 220 and/or c training data included in the area 230, is the same kind of training data as the training data of the target class (second training In the case of data), the generating device may classify learning data (b learning data and/or c learning data) of a target class corresponding to the nearest learning data as normal learning data.

In this specification, learning data in which all of the nearest learning data are heterogeneous learning data, such as a learning data, among the learning data, will be referred to as'isolated learning data'. In addition, in this specification, learning data in which at least one of the nearest learning data, such as b learning data and c learning data, is the same kind of learning data as the learning data, will be referred to as'normal learning data'.

The first training data and the second training data shown in FIG. 2 may be, for example, training data for machine learning or training of a neural network, or training data for a classifier of a neural network. The classifier can have a very close relationship with the mutual distribution between training data. Training data can exist densely or sparse. In addition, neighboring learning data adjacent to the learning data may have only the corresponding learning data and different types of learning data, or there may be a lot of learning data of the same kind as the corresponding learning data.

In an embodiment, a problem in which the performance of a classifier is biased may be solved by appropriately adjusting the number of training data or a ratio between training data by utilizing the distribution characteristic between the training data. In addition, in an embodiment, a lot of new learning data may be generated by analyzing the distribution characteristics of the learning data and surrounding situations, and giving weights to important learning data.

3 is a diagram for describing a method of generating additional learning data corresponding to isolation learning data according to an embodiment. 3, 6 additional training data oversampled from the isolated training data a1 (a _{1_1} , a _{1_2} , a _{1_3} , a _{1_4} , a _{1_5} , a _{1_6} ) and the isolated training data a1 Distances of the training data (da _{1 _1} , da _{1 _2} , da _{1 _3} , da _{1_4} , da _{1_5} , da _{1_6} ) are shown.

At this time, 6 additional training data (a _{1_1} , a _{1_2} , a _{1_3} , a _{1_4} , a _{1_5} , a _{1_ 6} ) indicated as new second training data in FIG. 3 are, for example, in the isolated training data a1 It can be generated by selecting the six closest learning data and selecting (or oversampling) an arbitrary point within a linear distance between the isolated learning data a1 and the six learning data as new additional learning data.

For instance, numbered, the number of of the entire training data oversampling learning data to (second learning data) N, the number of isolation of the N learning data, learning data is _one N, the number of the normal learning data N Let's say _two. In addition, suppose that the number of additional training data to be oversampled from N training data is S, and the number of _{additional training data extracted from N 1 isolated training data is S 1.}

_{The generating device is based on the first number (S 1} ) of the additional training data to be oversampled from the isolated training data (a1), from the training data of the N target classes (second training data) to a third number (for example, , S ₁ /N ₁ ) of (closest) additional training data can be extracted.

The generating device may _{calculate a distance (first distance) between N 1} pieces of isolated training data and a third number (S ₁ /N ₁ piece) of (closest) additional training data. Here, the distance is a distance representing the similarity between the isolation learning data and the additional learning data extracted from the isolation learning data, and may correspond to a Euclidean distance such as L1-norm (norm) or L2-norm. The first distance may be, for example, da _{1 _1} , da _{1 _2} , da _{1 _3} , da _{1 _4} , da _{1 _5} , and da _{1 _6} .

)를 산출할 수 있다. The generation device based on the first distance, the number of samples per unit distance of the isolated learning data (

) Can be calculated.

The generation device may calculate the number of sampling per unit distance of the isolated learning data by, for example, [Equation 1] below.

은 최근접 학습 데이터들 간의 거리의 전체 합을 나타낸다.

은 예를 들어, 격리 학습 데이터(a1)와 격리 학습 데이터(a1)로부터 추출한 추가 학습 데이터들(a_{1_1}, a_{1_2}, a_{1_3}, a_{1_4}, a_{1_5}, a_{1_6}) 간의 거리(da_{1 _1}, da_{1 _2}, da_{1_3}, da_{1 _4}, da_{1 _5}, da_{1 _ 6})를 전부 합한 값에 해당할 수 있다. S1은 격리 학습 데이터로부터 오버샘플링할 추가 학습 데이터들의 제1 개수(S₁)에 해당할 수 있다. here,

Represents the total sum of the distances between the nearest learning data.

silver For example, the distance between the isolated learning data a1 and the additional training data extracted from the isolated learning data a1 (a _{1_1} , a _{1_2} , a _{1_3} , a _{1_4} , a _{1_5} , a _{1_6 ) (da 1 _1} , da _{1 _2} , da _{1_3} , da _{1 _4} , da _{1 _5} , da _{1 _ 6} ) may be the sum of all values. S1 may correspond to _{the first number (S 1} ) of additional training data to be oversampled from the isolated training data.

The generating device may oversample additional learning data corresponding to the isolated learning data in inverse proportion to the first distance, for example. The generating device can reduce the number of additional training data to be oversampled as the distance between the two training data (isolated training data and additional training data) increases, and increase the number of additional training data to be oversampled as the distance between the two training data is close. have.

For example, when the first distance is farther than the unit distance of the isolated learning data, the generating device may oversample the number of additional training data less than the number of samples per unit distance of the isolated learning data from the isolated learning data. In addition, when the first distance is equal to or close to the unit distance of the isolated learning data, the generating device may oversample the number of additional training data from the isolated learning data that is greater than the number of samples per unit distance of the isolated learning data. That is, the generating device may increase or decrease the number of additional training data according to the distance between the training data.

Since the isolated learning data exists far from the same kind of learning data, the generating device may increase the reliability of the generated additional learning data by generating more additional learning data by giving a weight to the learning data close to the isolated learning data.

FIG. 4 is a diagram for describing a method of removing additional learning data according to whether the additional learning data generated in response to the isolated learning data and the learning data are overlapped according to an embodiment. Referring to FIG. 4, a situation in which newly generated additional learning data (eg, second training data) is very close to heterogeneous training data (eg, first training data) is illustrated.

The generating device according to an embodiment may determine whether additional learning data corresponding to the isolation learning data and heterogeneous learning data overlap. The generating device, for example, calculates a distance between newly generated additional learning data (second learning data) and heterogeneous learning data (first learning data), and if the distance is less than or equal to a specific value (for example, r) It can be judged by overlapping.

The generating device may remove additional learning data corresponding to the isolated learning data based on a result of determining whether or not to overlap. When the overlapping occurs, the generating device may remove the overlapping additional learning data and generate new additional learning data corresponding to the isolated learning data.

5 is a diagram for describing a method of generating additional learning data corresponding to normal learning data according to an embodiment. 5, the distance between the normal training data (a2) and the additional training data (a _{2_1} , a _{2_2} , a _{2_3} , a _{2_4} , a _{2_5 ) generated from the normal training data (a2) (da 2 _1} , da _{2 _2} , da _{2 _3} , da _{2 _4} , da _{2 _5} ), and the distance between the normal training data (a3) and the additional training data (a _{3_1} , a _{3_2 ) generated from the normal training data (a3) (da 3 _1} , da _{3_2} ) is shown.

For example, assume that the number of total training data is N, of which the number of isolated training data is N ₁ and the number of normal training data is N ₂ . In addition, suppose that the number of additional training data to be additionally generated from N training data is S, and the number of additional training data extracted from _{N 2 normal training data is S 2.}

_{The generation device is based on the second number (S 2} ) of additional training data to be oversampled from the normal training data, based on the fourth number (for example, from the N total training data (isolated training data + normal training data)). S ₂ /N ₂ ) (closest) additional learning data can be extracted.

The generating device may _{calculate a distance (second distance) between N 2} pieces of normal training data and a fourth number (S ₂ /N ₂ pieces) of (closest) additional training data. The second distance may be, for example, da _{2_1} , da _{2_2} , da _{2_3} , da _{2_4} , da _{2_5,} and da _{3_1} , da _{3_2} .

)를 산출할 수 있다. The generation device based on the second distance, the number of samples per unit distance of the normal learning data (

) Can be calculated.

The generation device may calculate the number of samplings per unit distance of the normal learning data by, for example, [Equation 2] below.

은 최근접 학습 데이터들의 거리의 전체 합을 나타내고, T 는 학습 데이터들 각각의 최근접 학습 데이터들의 개수이다.

은 예를 들어, 노말 학습 데이터들(a2, a3)과 노말 학습 데이터들로부터 추출한 추가 학습 데이터들(a_{2_1}, a_{2_2}, a_{2_3}, a_{2_4}, a_{2_5}, 및 a_{3_1}, a_{3_2}) 간의 거리(da_{2 _1}, da_{2 _2}, da_{2 _3}, da_{2 _4}, da_{2_5} 및 da_{3_1}, da_{3_2})를 전부 합한 값에 해당할 수 있다. here

Denotes the total sum of the distances of the nearest training data, and T is the number of the nearest training data of each of the training data.

Is, for example, the distance between the normal training data (a2, a3) and additional training data extracted from the normal training data (a _{2_1} , a _{2_2} , a _{2_3} , a _{2_4} , a _{2_5} , and a _{3_1} , a _{3_2)} It may correspond to the sum of all of (da _{2 _1} , da _{2 _2} , da _{2 _3} , da _{2 _4} , da _{2_5} and da _{3_1} , da _{3_2 ).}

Since oversampling is performed between training data (second training data) of the target class, T may vary for each training data.

The generating device may oversample additional training data corresponding to the normal training data in proportion to the second distance. The generation device may increase the number of additional training data to be oversampled as the distance between the two training data increases, and may decrease the number of additional training data to be oversampled as the distance increases.

For example, when the second distance is farther than the unit distance of the normal training data, the generating device may oversample the number of additional training data more than the number of samples per unit distance of the normal training data from the normal training data. In addition, when the second distance is equal to or close to the unit distance of the normal training data, the generating device may oversample the number of additional training data less than the number of samples per unit distance of the normal training data from the normal training data.

Since the normal learning data exists closer to the heterogeneous learning data than the isolated learning data, the generating device weights the learning data farther from the normal learning data to generate more learning data, so the reliability of the additional learning data generated. Can improve. In this case, the generation of a lot of additional training data in the training data located adjacent to the normal training data is not very helpful in improving the performance of the classifier compared to an increase in quantity.

6 is a flowchart illustrating a method of generating learning data according to another exemplary embodiment. Referring to FIG. 6, the generation device according to an embodiment performs oversampling from the training data of the target class and the number (N) of training data of the target class to be oversampled among training data for different classes. When the number of additional training data S is input (605 ), the generating device may extract S/N nearest learning data for each of the training data of the target class (610 ).

The generating device may calculate the number of isolated learning data among the learning data of the target class by classifying them into isolation learning data and normal learning data based on the nearest learning data for each of the learning data of the target class (615). . In this case, the generating device may also calculate the number of normal training data among the training data of the target class.

The generating device may calculate a ratio between the isolation learning data and the normal learning data based on the number of isolation learning data and the number of normal haze data calculated in step 615. For example, let's say the number of isolated learning data is S1, and normal learning data is S2. In addition, it is assumed that the number of isolated training data among the N training data of the target class to be oversampled is N1 and the number of normal training data is N2. The generating device may generate additional learning data corresponding to each of the isolated learning data and the normal learning data by dividing the isolated learning data and the normal learning data.

For the S1 isolated learning data, the generating device may extract S1/N1 nearest additional learning data from the training data of the target class to be oversampled (625).

The generating device may calculate a distance (first distance) between the isolation learning data and the additional learning data extracted from the isolation learning data (630), and calculate the number of samples per unit distance of the isolation learning data based on the first distance. There is (635).

The generating device may sample additional learning data corresponding to the isolated learning data in inverse proportion to the first distance (640 ). According to an embodiment, the generating device may check whether additional learning data corresponding to the isolated learning data generated through operation 640 and heterogeneous learning data are overlapped (645 ). The generating device may remove the overlapped additional learning data, and may newly oversample the additional learning data corresponding to the isolated learning data in place of the removed additional learning data.

In addition, for S2 normal training data, the generating device may extract S2/N2 nearest additional training data from the entire training data (650).

The generating device may calculate a distance (second distance) between the normal learning data and the additional learning data extracted from the normal learning data (655), and calculate the number of samples per unit distance of the isolated learning data based on the second distance. There is (660).

The generating device may sample additional learning data corresponding to the normal learning data in proportion to the second distance (665 ).

The generating device may combine the additional learning data sampled through steps 640 and 665 and use them as final additional learning data (670 ).

In one embodiment, learning data with high reliability is generated by using all distributions of not only the same kind of learning data but also the heterogeneous learning data, and generating new additional learning data based on appropriate weights within the isolated learning data and the normal learning data. Can be generated.

7 is a block diagram of an apparatus for generating training data according to an exemplary embodiment. Referring to FIG. 7, an apparatus for generating training data (hereinafter, a generation apparatus) 700 according to an exemplary embodiment includes a processor 710. The generating device 700 may further include a communication interface 730 and a memory 750. The processor 710, the communication interface 730, and the memory 750 may communicate with each other through a communication bus 705. The generation device 700 may be, for example, an artificial intelligence classifier for machine learning or a classifier for a neural network, or may include a classifier.

The processor 710 divides training data to be oversampled into isolation training data and normal training data based on distribution characteristics of the training data. The processor 710 determines the number of additional training data to be oversampled from each of the isolated training data and the normal training data according to a ratio between the isolated training data and the normal training data. The processor 710 generates additional training data corresponding to each of the isolated training data and the normal training data, based on the number of each of the training data to be oversampled.

The processor 710 includes training data of the target class based on a ratio between training data of the target class to be oversampled and additional training data to be oversampled from the training data of the target class among training data for different classes. The nearest learning data for each can be extracted.

The processor 710 may classify the training data of the target class into isolation training data and normal training data based on the nearest training data for each of the training data of the target class.

The processor 710 may extract closest training data corresponding to a ratio between training data of the target class and additional training data to be oversampled based on a distance between each of the training data of the target class.

The processor 710 may classify the learning data of the target class corresponding to the nearest learning data as isolation learning data when all of the nearest learning data are learning data of a target class and different types of learning data. When at least one of the closest training data is training data of the same kind as the training data of the target class, the processor 710 may classify training data of the target class corresponding to the closest training data as normal training data.

The processor 710 may determine a first number of additional training data to be oversampled from the isolated training data and a second number of additional training data to be oversampled from the normal training data according to a ratio between the isolated training data and the normal training data. .

The processor 710 may extract additional training data based on the first number of additional training data to be oversampled from the isolated training data and the second number of additional training data to be oversampled from the normal training data. The processor 710 is based on a first distance between the isolated learning data and the additional learning data extracted from the isolated learning data, and a second distance between the normal learning data and the additional learning data extracted from the normal learning data. Additional training data corresponding to each of the training data may be generated.

In addition, the processor 710 may perform at least one method described above with reference to FIGS. 1 to 6. The processor 710 may execute a program and control the generating device 700. The program code executed by the processor 710 may be stored in the memory 750.

The communication interface 730 may receive training data from outside of the generating device 700. In addition, the communication interface 730 may transmit additional learning data generated by the processor 710 to the outside of the generating device 700.

The memory 750 may store training data received from the communication interface 730. The memory 750 may store additional learning data generated by the processor 710. In addition, the memory 750 may store various types of information generated during processing in the processor 710 described above. In addition, the memory 750 may store various types of data and programs. The memory 750 may include a volatile memory or a nonvolatile memory. The memory 750 may include a mass storage medium such as a hard disk to store various types of data.

The embodiments described above may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component. For example, the devices, methods, and components described in the embodiments are, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate (FPGA). array), programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications executed on the operating system. Further, the processing device may access, store, manipulate, process, and generate data in response to the execution of software. For the convenience of understanding, although it is sometimes described that one processing device is used, one of ordinary skill in the art, the processing device is a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of these, configuring the processing unit to operate as desired or processed independently or collectively. You can command the device. Software and/or data may be interpreted by a processing device or, to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. , Or may be permanently or temporarily embodyed in a transmitted signal wave. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

As described above, although the embodiments have been described by the limited drawings, a person of ordinary skill in the art can apply various technical modifications and variations based on the above. For example, the described techniques are performed in a different order from the described method, and/or components such as systems, structures, devices, circuits, etc. described are combined or combined in a form different from the described method, or other components Alternatively, even if substituted or substituted by an equivalent, an appropriate result can be achieved.

700: generating device
705: communication bus
710: processor
730: communication interface
750: memory

Claims (20)

In the method for generating training data by the generating device,
Dividing, by the generating device, training data to be oversampled into isolation training data and normal training data based on distribution characteristics of the training data;
Determining, by the generating device, the number of additional learning data to be oversampled from each of the isolated learning data and the normal learning data according to a ratio between the isolated learning data and the normal learning data; And
Generating, by the generating device, additional learning data corresponding to each of the isolation learning data and the normal learning data, based on the number of each of the learning data to be oversampled.
Including,
Generating the additional learning data,
The generating device comprises: a first number of additional learning data to be oversampled from the isolated learning data determined according to a ratio between the isolated learning data and the normal learning data and a second number of additional learning data to be oversampled from the normal learning data On the basis of, extracting the additional learning data; And
The generating device, based on each of a first distance between the isolation learning data and additional learning data extracted from the isolation learning data and a second distance between the normal learning data and additional learning data extracted from the normal learning data, Generating additional learning data corresponding to each of the isolation learning data and the normal learning data
Containing, a method of generating training data. The method of claim 1,
The step of distinguishing
The generation device, based on a ratio between training data of a target class to be oversampled among the training data for different classes and additional training data to be oversampled from training data of the target class, the Extracting nearest learning data for each of the training data of the target class; And
Dividing, by the generating device, the learning data of the target class into the isolation learning data and the normal learning data, based on the nearest learning data for each of the learning data of the target class
Containing, a method of generating training data. The method of claim 2,
Extracting the nearest learning data for each of the training data of the target class is
Extracting, by the generating device, closest learning data corresponding to a ratio between learning data of the target class and additional learning data to be oversampled based on a distance between each of the training data of the target class
Containing, a method of generating training data. The method of claim 2,
Based on the nearest learning data, the step of dividing the learning data of the target class into the isolation learning data and the normal learning data
When all of the nearest learning data are learning data of the target class and learning data of the target class, dividing learning data of a target class corresponding to the nearest learning data into the isolation learning data ; And
When at least one of the nearest learning data is learning data of the same kind as the learning data of the target class, the generating device divides learning data of a target class corresponding to the nearest learning data into the normal learning data. Steps to do
Containing, a method of generating training data. delete delete The method of claim 1,
Extracting the additional learning data
Extracting, by the generating device, a third number of additional training data from training data of a target class based on a first number of additional training data to be oversampled from the isolated training data; And
Extracting, by the generating device, a fourth number of additional training data from the training data based on a second number of additional training data to be oversampled from the normal training data
Containing, a method of generating training data. The method of claim 7,
The step of generating the additional learning data
The generating device, based on each of a first distance between the isolation learning data and the third number of additional learning data and a second distance between the normal learning data and the fourth number of additional learning data, the isolation learning Calculating the number of samplings per unit distance of the data and the normal learning data; And
The generating device, based on the first distance, the second distance, and the number of samples per unit distance of each of the isolation learning data and the normal learning data, additional learning corresponding to each of the isolation learning data and the normal learning data Steps to generate data
Containing, a method of generating training data. The method of claim 8,
Calculating the number of samplings per unit distance of each of the isolated learning data and the normal learning data
Calculating, by the generating device, a first distance between the isolation learning data and the third number of additional learning data; And
Calculating, by the generating device, a sampling number per unit distance of the isolated learning data based on the first distance
Containing, a method of generating training data. The method of claim 8,
The step of generating the additional learning data
Oversampling, by the generating device, additional learning data corresponding to the isolated learning data in inverse proportion to the first distance
Containing, a method of generating training data. The method of claim 10,
Oversampling additional learning data corresponding to the isolated learning data
When the first distance is greater than a unit distance of the isolated learning data, oversampling, by the generating device, a number of additional training data less than the number of samples per unit distance of the isolated learning data from the isolated learning data; And
When the first distance is equal to or close to the unit distance of the isolation learning data, the generating device oversampling a number of additional learning data greater than the number of samples per unit distance of the isolation learning data from the isolation learning data
Containing, a method of generating training data. The method of claim 1,
The step of generating the additional learning data
Determining, by the generating device, whether additional learning data corresponding to the isolated learning data and heterogeneous learning data overlap; And
Removing, by the generating device, additional learning data corresponding to the isolation learning data based on a result of the determination as to whether or not to overlap
Further comprising, a method of generating training data. The method of claim 8,
Calculating the number of samplings per unit distance of each of the isolated learning data and the normal learning data
Calculating, by the generating device, a second distance between the normal learning data and the fourth number of additional learning data; And
Calculating, by the generating device, the number of samplings per unit distance of the normal learning data based on the second distance
Containing, a method of generating training data. The method of claim 13,
The step of generating the additional learning data
Oversampling, by the generating device, additional learning data corresponding to the normal learning data in proportion to the second distance
Containing, a method of generating training data. The method of claim 14,
The step of oversampling additional training data corresponding to the normal training data
If the second distance is greater than a unit distance of the normal training data, the generating device oversampling a number of additional training data greater than the number of samples per unit distance of the normal training data from the normal training data; And
If the second distance is equal to or close to the unit distance of the normal learning data, the generating device oversampling a number of additional learning data less than the number of samples per unit distance of the normal learning data from the normal learning data
Containing, a method of generating training data. A computer program stored in a medium for executing the method of any one of claims 1 to 4 and 7 to 15 in combination with hardware. Based on the distribution characteristics of the training data, the training data to be oversampled are divided into isolated training data and normal training data,
Determine the number of additional learning data to be oversampled from each of the isolation learning data and the normal learning data according to a ratio between the isolation learning data and the normal learning data,
Generating additional learning data corresponding to each of the isolation learning data and the normal learning data based on the number of each of the learning data to be oversampled
Processor
Including,
The processor,
Based on a first number of additional learning data to be oversampled from the isolated learning data determined according to a ratio between the isolated learning data and the normal learning data and a second number of additional learning data to be oversampled from the normal learning data, the Extract additional learning data,
Based on a first distance between the isolation learning data and additional learning data extracted from the isolation learning data, and a second distance between the normal learning data and additional learning data extracted from the normal learning data, the isolation learning data and the An apparatus for generating training data that generates additional training data corresponding to each of the normal training data. The method of claim 17,
The processor is
Based on the ratio between the training data of the target class to be oversampled among the training data for different classes and the additional training data to be oversampled from the training data of the target class, each of the training data of the target class Extract the nearest learning data for
The apparatus for generating learning data, dividing the learning data of the target class into the isolation learning data and the normal learning data based on nearest learning data for each of the learning data of the target class. The method of claim 18,
The processor is
An apparatus for generating learning data, which extracts nearest learning data corresponding to a ratio between the learning data of the target class and the additional learning data to be oversampled, based on the distance between each of the training data of the target class . The method of claim 19,
The processor is
When all of the nearest learning data are learning data of the target class and learning data of a different type, the learning data of a target class corresponding to the nearest learning data is divided into the isolation learning data,
When at least one of the nearest learning data is training data of the same kind as the training data of the target class, training data for dividing training data of a target class corresponding to the nearest training data into the normal training data Generating device.

Images