KR102201201B1 - Method, computer program for classifying data and apparatus using the same - Google Patents

Abstract

A data classification apparatus for classifying data, comprising: a distribution map generator that generates a distribution map for each of a plurality of classes; and a search split point setting that sets a search split point that becomes a reference point for discretizing continuous variables included in a data set. Second, a final dividing point determination unit for determining a final dividing point based on a distance value derived by a preset distance function for the left area and the right area based on the search dividing point, and the continuation based on the final dividing point A discretization execution unit that obtains a categorical variable by performing discretization of a type variable, and data for classifying the data set into one of a plurality of preset classes by inputting at least a part of the data set and the categorical variable into a machine learning-based model It may include a classification unit.

Description

A method, a program for classifying data, and a device using the same {METHOD, COMPUTER PROGRAM FOR CLASSIFYING DATA AND APPARATUS USING THE SAME}

The present invention relates to a method, a program for classifying data, and an apparatus using the same.

Data variables are divided into discrete variables and continuous variables according to their characteristics. Continuous variables included in a data set contain a lot of information about the data. However, discrete variables are more concise and more convenient to use than continuous variables in determining specific reference points, such as determining whether or not obesity from the BMI index or assigning grades of grades based on a certain score. For this reason, continuous variables are converted to discrete variables.

There are classification algorithms used for data with a mixture of discrete and continuous variables, but certain classification algorithms require nominal variables. Therefore, in order to use the algorithm, a continuous variable must be changed to a discrete variable, and such a data processing process is called discretization.

On the other hand, in the process of discretizing the continuous variable, much of the information that the continuous variable had previously was lost, but the advantages of the discrete variable are as follows.

First, the speed of the classification algorithm is improved. For example, continuous variables have different values per record and different ranges of variable values for each variable. Therefore, the amount of calculation increases in the process of calculating with other variables. In addition, discretization reduces the load because only a small amount of variable values is calculated.

Second, the data becomes concise. For example, the number of discrete variables can be determined by the user, and usually a small number is defined and used. Therefore, the role of the discrete variable value is only to indicate the inclusion category of the data value, which is convenient for the user to obtain information from the data and to utilize it.

Third, when an appropriate discretization method is selected in consideration of the data set and classification algorithm, the performance of the classification algorithm can be expected to improve. For example, if the MDLP discretization technique using the amount of information acquired and C4.5 using the amount of information acquired are used together, it shows excellent performance through positive interaction.

As described above, research related to discretization has been conducted for a long time, and research is still being conducted in various fields. Among them, the most basic discretization techniques include equal width discretization and equal frequency discretization, which are called Binning techniques. This is an unsupervised discrete technique that does not refer to a class value, and because it is divided using a certain width and a certain frequency, there is little expectation for performance improvement. Another technique that does not refer to class values, the K-means clustering discretization technique, is also difficult to expect a noticeable performance improvement.

In addition, the conventional clustering-based discretization has a problem that the result of the discretization may vary each time because a different cluster is formed for each experiment.

The present invention is to solve the above-described problems of the prior art, and set a dividing point based on the intersection of the distribution map for each of a plurality of classes, but based on the distance values for the left area and the right area based on the set dividing point The position of the dividing point is changed, and the point where the sum of the distance value for the left area and the distance value for the right area is the minimum based on the dividing point at the changed position is determined as the final dividing point. An object of the present invention is to provide a data classification apparatus that classifies data that improves the accuracy of the discretization result and enables faster discretization by performing discretization of a continuous variable as a reference, a data classification method and a computer program using the same.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problems as described above, and other technical problems may exist.

As a technical means for achieving the above-described technical problem, an embodiment of the present invention relates to a data classification device for classifying data, a distribution map generator for generating a distribution map for each of a plurality of classes, and included in a data set. A search split point setting unit that sets a search split point that becomes a reference point for discretizing a continuous variable, and a final result based on a distance value derived by a preset distance function for the left area and the right area based on the search split point. A final splitting point determining unit for determining a splitting point, a discretization performing unit for obtaining a categorical variable by performing discretization of the continuous variable based on the final splitting point, and at least a part of the dataset and the categorical variable It may include a data classifier for inputting into a learning-based model and classifying the data set into one of a plurality of preset classes.

Another embodiment of the present invention relates to a data classification method for classifying data, the step of generating a distribution map for each of a plurality of classes, a search split point serving as a reference point for discretizing a continuous variable included in a data set The step of setting, determining a final dividing point based on a distance value derived by a preset distance function for a left area and a right area based on the search dividing point, the continuous type based on the final dividing point Performing discretization of variables to obtain categorical variables, and classifying the data set into one of a plurality of preset classes by inputting at least a portion of the data set and the categorical variable into a machine learning-based model. I can.

Another embodiment of the present invention, when executed by a computing device as a computer program, generates a distribution map for each of a plurality of classes, and provides a search splitting point as a reference point for discretizing a continuous variable included in a data set. Set and determine a final dividing point based on a distance value derived by a preset distance function for the left area and the right area based on the search dividing point, and discretize the continuous variable based on the final dividing point A medium comprising a sequence of instructions for acquiring a categorical variable by performing, and inputting at least a part of the data set and the categorical variable into a machine learning-based model to classify the data set into one of a plurality of preset classes Computer programs stored in can be provided.

The above-described problem solving means are merely exemplary and should not be construed as limiting the present invention. In addition to the above-described exemplary embodiments, there may be additional embodiments described in the drawings and detailed description of the invention.

According to any one of the above-described problem solving means of the present invention, the present invention sets the dividing point based on the intersection point of the distribution map for each of the plurality of classes, but the distance values for the left area and the right area based on the set dividing point The position of the dividing point is changed based on the dividing point, and the point at which the sum of the distance value for the left area and the distance value for the right area is the minimum is determined as the final dividing point based on the divided point at the changed position. By performing discretization of a continuous variable based on a point, it is possible to provide a data classification apparatus, a data classification method and a computer program using the same, for classifying data that enables faster discretization by searching for a split point more quickly.

In addition, the present invention determines the point at which the sum of the distance value for the left area and the distance value for the right area is the minimum based on the dividing point as the final dividing point, and discretizes the continuous variable based on the final dividing point. By performing, it is possible to provide a data classification apparatus for classifying data that discretizes a continuous variable capable of deriving more accurate discretization results, a data classification method and a computer program using the same.

1 is a diagram showing the configuration of a data classification apparatus used in a machine learning-based data classifier according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating setting of a search split point in a search split point setting unit according to an embodiment of the present invention.
FIG. 3 is a diagram illustrating determining a final dividing point by a final dividing point determiner according to an embodiment of the present invention.
4 is a diagram showing a process of discretizing a continuous variable according to an embodiment of the present invention.
5 is a diagram showing an experiment result of the accuracy of a discretization technique using a data classification apparatus used in a machine learning-based data classifier according to an embodiment of the present invention.
6 is a diagram showing an experiment result of an AUC of a discretization technique using a data classification apparatus used in a machine learning-based data classifier according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and similar reference numerals are assigned to similar parts throughout the specification.

Throughout the specification, when a part is said to be "connected" to another part, this includes not only "directly connected" but also "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, it means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

In the present specification, the term "unit" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. Further, one unit may be realized using two or more hardware, or two or more units may be realized using one hardware.

In the present specification, some of the operations or functions described as being performed by the terminal or device may be performed instead by a server connected to the terminal or device. Likewise, some of the operations or functions described as being performed by the server may also be performed by a terminal or device connected to the server.

Hereinafter, with reference to the accompanying configuration diagram or processing flow chart, specific details for the implementation of the present invention will be described.

1 is a diagram showing the configuration of a data classification apparatus 100 used in a machine learning-based data classifier according to an embodiment of the present invention.

Referring to FIG. 1, a data classification apparatus 100 used in a machine learning-based data classifier includes a distribution map generator 110, a search split point setting unit 130, a final split point determination unit 150, and a discretization execution unit ( 170) and a data classification unit 190 may be included.

The distribution map generator 110 may generate a distribution map for each of a plurality of classes.

It is to be noted that although it may be desirable for the distribution map generator 110 to generate a distribution map for each of at least two classes, it is not limited thereto.

The distribution map generator 110 may generate a distribution map in advance based on initial data for each of a plurality of classes.

The search split point setting unit 130 may set a search split point that is a reference point for discretizing a continuous variable included in the data set.

Also, the search split point setting unit 130 may set the first search split point based on the intersection points of a plurality of distributions included in the distribution map.

In addition, the search split point setting unit 130 may change the location of the search split point. Here, the search split point setting unit 130 may repeatedly change the position of the search split point.

The final split point determiner 150 may determine the final split point based on a distance value derived by a preset distance function for the left and right areas of the distribution map based on the search split point.

Here, the distance function is a Wasserstein distance function, and may be defined based on Equation 1 below.

(Where P, Q are probability distributions, x, y are random variables)

Also, the final split point determiner 150 may calculate a sum of a first distance value for the left area and a second distance value for the right area based on the search split point at the changed location. Here, the final split point determination unit 150 may repeat the calculation of the sum of the first distance value and the second distance value whenever the search split point setting unit 130 repeatedly changes the position of the search split point.

Here, the first distance value may be derived by Equation 2 below.

은 탐색 분할점을 기준으로 좌측에 있는 변수 값)(here,

Is the value of the variable on the left of the search split point)

In addition, the second distance value may be derived by Equation 3 below.

는 탐색 분할점을 기준으로 우측에 있는 변수 값)(here,

Is the value of the variable to the right of the search split point)

In addition, the final split point determiner 150 may determine a search split point in which the sum of the first distance value and the second distance value is the minimum as the final split point.

Here, when the sum of the first distance value and the second distance value is the minimum, it may mean that the distances between the different distribution maps are close, which may mean that there are many common parts between the different distributions. Note that although it may mean that the cost value transferred between distributions is small, it is not limited thereto.

The discretization performing unit 170 may perform discretization of the continuous variable based on the final dividing point.

In addition, the discretization performing unit 170 may obtain a categorical variable by performing discretization of a continuous variable.

Here, it should be noted that the continuous variable subject to discretization may mean a variable classified as having characteristics based on importance among continuous variables included in the data set, but is not limited thereto.

The data classification unit 190 may classify the data set into one of a plurality of preset classes by inputting at least a part of the data set and categorical variables into the machine learning-based model.

Here, at least a part of the dataset may mean data excluding the continuous variable discretized by the above-described discretization execution unit 170. In addition, the machine learning-based model is a decision tree, random forest, SVM, and K. It should be noted that it may be formed of at least one or more of -nn, but is not limited thereto.

Hereinafter, with reference to FIG. 2, the search split point setting unit 130 according to an embodiment of the present invention will be described in more detail.

FIG. 2 is a diagram for explaining setting of a search split point by the search split point setting unit 130 according to an embodiment of the present invention.

First, referring to FIG. 2A, the search split point setting unit 130 may set a search split point 250 that is a reference point for discretizing a continuous variable included in a data set.

Here, the search split point 250 may be the first search split point 250 set based on the intersection of a plurality of distributions included in the distribution map.

For example, the search split point setting unit 130 may set the first search split point 250 based on the intersection of distributions for each of the different classes 210 and 230.

The search split point setting unit 130 may determine an optimal split point for discretizing the continuous variable while changing the position of the first search split point 250.

FIG. 2B shows that the location of the search split point 270 is moved to the left from the location of the first search split point 250.

The final split point determiner 150 may determine the final split point based on a distance value derived by a preset distance function for the left and right areas of the distribution map based on the search split point.

For example, the final split point determiner 150 may calculate a sum of a first distance value for the left area and a second distance value for the right area based on the first search split point 250.

Also, the final split point determiner 150 may calculate a sum of the first distance value for the left area and the second distance value for the right area based on the search split point 270 at the changed position.

The final split point determiner 150 may determine a search split point having a minimum sum of the first distance value and the second distance value as the final split point.

That is, the search split point setting unit 130 may move the location of the search split point 250 initially set based on the intersection of a plurality of distributions included in the distribution map, in this case, the final split point determination unit ( 150) calculates the sum of the first distance value for the left area and the second distance value for the right area based on the search split point at the changed location, and the sum of the plurality of moved search split points is the minimum. The in search splitting point can be determined as the final splitting point.

On the other hand, the search split point setting unit 130 determines the first search split point 250 based on the nearest intersection point with the average value of all variable values when there are multiple intersections of the distributions of different classes 210 and 230. Can be set.

Hereinafter, with reference to FIG. 3, the final split point determination unit 150 according to an embodiment of the present invention will be described in more detail.

FIG. 3 is a diagram illustrating the determination of a final dividing point by the final dividing point determiner 150 according to an embodiment of the present invention.

Referring to FIG. 3A, the final splitting point determination unit 150 shows a distribution diagram based on the first search splitting point 250 set based on the intersection of distributions for each of the different classes 210 and 230. A final dividing point may be determined based on a distance value derived by a preset distance function for the left area 310 and the right area 330.

For example, the final dividing point determination unit 150 calculates the sum of the first distance value for the left area 310 and the second distance value for the right area 330 based on the first search dividing point 250. Can be calculated.

In addition, referring to FIG. 3B, the final divided point determining unit 150 is a left region 310 that is changed as the search divided point setting unit 130 repeats the change of the position of the search divided point 270. ) And the sum of the first distance value and the second distance value of the right region 330 can be repeated, and a search split point in which the sum of the first distance value and the second distance value derived through this process is the minimum It can be determined as the final split point.

Through this process, the data classification apparatus 100 of the present invention may derive more accurate discretization results.

The pseudo code for determining the final split point by the final split point determination unit 150 is as follows.

각각은 어느 하나의 클래스 값을 갖는 집합의 원소를 의미할 수 있고,

는 서로 다른 클래스 값을 갖는 데이터 셋의 합집합인 전체 데이터 셋을 의미할 수 있다.)(here,

Each can mean an element of a set having any one class value,

May mean the entire data set, which is the union of data sets having different class values.)

4 is a diagram showing a process of discretizing a continuous variable according to an embodiment of the present invention.

Referring to FIG. 4, in step S410, the distribution map generator 110 may generate a distribution map including a plurality of distributions for continuous variables included in each of a plurality of classes classified from a data set.

In step S430, the search split point setting unit 130 may set a search split point that is a reference point for discretizing a continuous variable included in the data set.

In step S450, the final split point determiner 150 may determine the final split point based on a distance value derived by a preset distance function for the left region of the distribution map based on the search split point.

In step S470, the discretization performing unit 170 may obtain a categorical variable by performing discretization of the continuous variable based on the final split point.

In step S490, the data classification unit 190 may classify the data set into one of a plurality of preset classes by inputting at least a part of the data set and categorical variables into the machine learning-based model.

In the above description, steps S410 to S490 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as necessary, and the order between steps may be changed.

Hereinafter, results of experiments on the classification performance of the data classification apparatus 100 according to an embodiment of the present invention will be described.

First, as for the data used in the experiment, all variables were continuous variables, there were no missing values, and only data sets having two class values were collected and used. Therefore, only the data set usable for the binary classification problem was obtained and used.

The data set was obtained from the UCI Machine Learning Repository, and information on the data set is shown in Table 1 below.

Data set Number of variables Number of records Sonar 60 208 Australian 14 690 Breast 9 699 Pima 8 768 Iono 32 351 Wdbc 30 569 heart 13 270

On the other hand, the device to be compared with the data classification device 100 according to an embodiment of the present invention is a device consisting of a discrete algorithm, MDLP (Minimum Description Length Principle) and a classification algorithm, and a discrete algorithm, CAIM (Complementary Alternative and Integrative Medicine), and classification. A device consisting of an algorithm, a device consisting of a discrete algorithm, CACC (Cooperative Adaptive Cruise Control) and a classification algorithm, a device consisting of a discrete algorithm AMEVA and a classification algorithm, is used, and a device consisting of the original data set and classification algorithm without discretization is also shown. Used for performance comparison. Here, as the classification algorithm, SVM and Random forest were used. Note that in FIGS. 5 and 6, each device to be compared is represented by a discrete algorithm.

Also, the above discretization method was used by calling the code written in the R library “Discretization” registered in CRAN. The comparison of classification performance was accepted using the Accuracy and Area under the ROC (AUC) scales.

Hereinafter, results of experiments on the classification performance of the data classification apparatus 100 according to an exemplary embodiment will be described with reference to FIGS. 5 and 6.

5 is a diagram showing results of an experiment on the classification accuracy of the data classification apparatus 100 according to an embodiment of the present invention.

6 is a diagram showing an experiment result of an area under the ROC (AUC) of the data classification apparatus 100 according to an embodiment of the present invention.

First, referring to FIG. 5, in the case of accuracy, the data classification apparatus 100 of the present invention exhibits superior performance with respect to a plurality of data sets than other devices.

Referring to FIG. 6, in the case of classification performance, the data classification apparatus 100 of the present invention exhibits the best performance for three data sets.

However, in the case of using the data classification apparatus 100 of the present invention, the accuracy performance for the sonar and iono data sets among the data sets was not good due to a slight difference compared to other devices, but this is a variable based on the characteristics of the sonar and iono data sets. It is judged to have occurred because the number of records was too small compared to the number.

That is, it can be seen that the data classification apparatus 100 of the present invention can derive a more accurate classification result.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer-readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Further, the computer-readable medium may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The above description of the present invention is for illustrative purposes only, and those of ordinary skill in the art to which the present invention pertains will be able to understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not limiting. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. .

100: data classification device
110: distribution map generator
130: search split point setting unit
150: final division point determination unit
170: discretization execution unit
190: data classification unit

Claims (14)

In the data classification device for classifying data,
A distribution map generator for generating a distribution map including a plurality of distributions for continuous variables included in each of the plurality of classes classified from the data set;
A search split point setting unit for setting a search split point as a reference point for discretizing the continuous variable included in the data set;
A final dividing point determination unit determining a final dividing point based on a distance value derived by a preset distance function for a left area and a right area of the distribution map based on the search dividing point;
A discretization performing unit for obtaining a categorical variable by performing discretization of the continuous variable based on the final splitting point; And
Data classification unit for classifying the data set into one of a plurality of preset classes by inputting at least a part of the data set and the categorical variable into a machine learning-based model
That includes, data classification device.
The method of claim 1,
The data classification apparatus, wherein the search split point setting unit sets an initial search split point based on an intersection of the plurality of distributions included in the distribution map.
The method of claim 2,
The search split point setting unit changes the position of the search split point,
The data classification apparatus, wherein the final split point determination unit calculates a sum of a first distance value for the left area and a second distance value for the right area based on the search split point at the changed position.
The method of claim 3,
The search split point setting unit repeats the change of the position of the search split point,
Wherein the final split point determination unit repeats calculation of the sum of the first distance value and the second distance value.
The method of claim 3,
Wherein the final splitting point determining unit determines a search splitting point having a minimum sum of the first distance value and the second distance value as the final splitting point.
The method of claim 1,
The preset distance function is defined by the following equation.

(Where P, Q are probability distributions, x, y are random variables)
The method of claim 3,
The data classification apparatus, wherein the first distance value is derived by the following equation.

The method of claim 3,
The second distance value is derived by the following equation.

In the data classification method for classifying data,
Generating a distribution map including a plurality of distributions for continuous variables included in each of the plurality of classes classified from the data set;
Setting a search splitting point as a reference point for discretizing the continuous variable included in the data set;
Determining a final dividing point based on a distance value derived by a preset distance function for a left area and a right area of the distribution map based on the search dividing point;
Obtaining a categorical variable by performing discretization of the continuous variable based on the final dividing point; And
Classifying the dataset into one of a plurality of preset classes by inputting at least a part of the dataset and the categorical variable into a machine learning-based model
That includes, data classification method.
The method of claim 9,
The step of setting the search split point includes setting an initial search split point based on the intersection of the plurality of distributions included in the distribution map.
The method of claim 10,
The step of setting the search dividing point includes changing a position of the search dividing point,
The determining of the final dividing point includes calculating a sum of a first distance value for the left area and a second distance value for the right area based on the search dividing point at the changed position. , Data classification method.
The method of claim 11,
The step of setting the search dividing point includes repeating the change of the position of the search dividing point,
The determining of the final dividing point includes repeating calculation of the sum of the first distance value and the second distance value.
The method of claim 11,
The determining of the final dividing point includes determining a search dividing point having a minimum sum of the first distance value and the second distance value as the final dividing point.
A computer program stored on a medium comprising a sequence of instructions for classifying data,
When the computer program is executed by a computing device,
Generate a distribution map including a plurality of distributions for continuous variables included in each of the plurality of classes classified from the dataset,
Set a search split point that becomes a reference point for discretizing the continuous variable included in the dataset,
A final dividing point is determined based on a distance value derived by a preset distance function for a left area and a right area of the distribution map based on the search dividing point,
To obtain a categorical variable by performing discretization of the continuous variable based on the final dividing point,
A computer program stored in a medium comprising a sequence of instructions for inputting at least a portion of the dataset and the categorical variable into a machine learning-based model to classify the dataset into one of a plurality of preset classes.

Images