KR101897513B1 - Method, apparatus and computer-readable recording medium for selecting feature using relational matrix - Google Patents

Abstract

The present invention relates to a method, a device, and a computer-readable recording medium for selecting a feature using a relational matrix, which generate a matrix of input data and count an association between the data by using a relational matrix, thereby finally selecting the most associated minority features. According to an embodiment of the present invention, the method for selecting a feature using a relational matrix comprises the following steps: generating the matrix composed of a row number and a column number set for each of the input data; counting the number of occurrences of the data appearing together with a specific event by using the generated matrix whenever the specific event occurs; selecting the data in which the number of occurrences is the highest among the input data; and extracting feature data associated with the selected data.

Description

[0001] METHOD, APPARATUS AND COMPUTER READABLE RECORDING MEDIUM FOR SELECTING FEATURE USING RELATIONAL MATRIX [0002]

The present invention relates to a feature selection method using a relational matrix and a computer-readable recording medium, and more particularly, to a feature selection method, apparatus, and computer readable medium using a relational matrix for selecting features based on correlation of input data Recording medium.

Of the approximately 30,000 genes in humans that have information, the number of genes associated with disease is only a fraction of the 30,000.

There are Sequential Backward Search methods such as Overlapped Area Measurement and Match Count method for extracting a very small number of characteristic data that are relevant among a large amount of data, such as finding a gene related to a disease among human genes.

However, since the sequential backward search method is a method of finding the most disadvantageous features and eliminating them one by one, it takes a long time to find an appropriate feature corresponding to a very small number of features.

In the past, algorithms such as Gini Index and Maximum Relevance and Minimum Redundancy (MRMR) were developed based on the correlation between features.

The Gini Index algorithm is an algorithm that finds the proper features by measuring how heterogeneous is mixed in the set. The features have Gini index calculated by binary classification, and the top k Features are selected as appropriate features.

The MRMR algorithm uses mutual information between variables to select a subset of features with a maximum relevancy while having a minimum redundancy. Select features.

The Gini Index and MRMR algorithms are useful for calculating correlations between features of similar characteristics, such as the width of two petals, the number of genes, and the number of wheels.

However, in the case of machine learning, since the similarity between features used as input variables is not always guaranteed, the development of an algorithm for more accurately estimating the correlation between input variables for which similarity is not guaranteed, need.

A Comparative Analysis of Feature Selection Algorithms on Classification of Gene Microarray Dataset, J. jeyachidra, M. punithavalli. February 21, 2013, Information Communication and Embedded Systems (ICICES), 2013 International Conference on.

The present invention relates to a feature selection method, a device and a computer readable recording method using a relational matrix for generating a matrix of input data, using a relational matrix to determine whether the data are related or not, and finally selecting the most- Media.

A feature selection method using a relational matrix according to an exemplary embodiment of the present invention includes generating a matrix including a row number and a column number set for each of input data; Counting the number of occurrences of data appearing together with the specific event using the generated matrix every time a specific event occurs; Selecting data in which the number of occurrences is the highest among the input data; And extracting feature data associated with the selected data.

The generating may include generating an n x n matrix when the number of input data is n.

Wherein the counting step includes the steps of: selecting row numbers set for data appearing together with the specific event among the row numbers of the matrix; Selecting column numbers set for data appearing together with the specific event among column numbers of the matrix; And counting the number of occurrences of the data while increasing values of elements included in the selected column numbers among the elements included in the selected row numbers in the matrix.

Wherein the step of selecting the data comprises: selecting a row number of an element having a largest value among elements included in a diagonal component of the matrix; And selecting data corresponding to the selected row number.

Wherein the extracting of the feature data comprises: selecting column numbers including N upper elements in order of magnitude among elements included in the selected row number of the matrix; And extracting feature data corresponding to the row numbers including the upper N elements in order of magnitude among the elements included in the selected column numbers for each selected column number of the matrix.

The extracting of the characteristic data may include generating sub-matrices in which the elements included in the respective columns are sorted in descending order of size for each selected column number of the matrix; Extracting elements included in the partial matrixes from the first row and the first column to the N row and the first column of each of the generated sub-matrices; And extracting feature data corresponding to row numbers including the elements extracted from the respective sub-matrices, in each of the column numbers corresponding to the sub-matrices among the column numbers of the matrix.

N is a natural number equal to or less than a square root of the number of extracted data and equal to or less than the number of the input data.

The submatrices are n x N matrices when the number of input data is n.

A feature selection apparatus using a relational matrix according to an exemplary embodiment of the present invention generates a matrix including row numbers and column numbers set for each of input data, and generates a matrix using the generated matrix It is possible to count the number of occurrences of data appearing together with a specific event, select data in which the number of occurrences is the highest among the input data, and extract the feature data associated with the selected data.

The apparatus may generate a matrix of size n x n if the number of input data is n.

Wherein the apparatus selects row numbers set for data appearing together with the specific event among the row numbers of the matrix and sets a column number set for data appearing together with the specific event among the column numbers of the matrix The number of occurrences of the data can be counted while increasing the value of the elements included in the selected column numbers among the elements included in the selected row numbers in the matrix.

The apparatus can select the row number of the element having the largest value among the elements included in the diagonal component of the matrix and select the data corresponding to the selected row number.

The apparatus includes column numbers including upper N elements in order of magnitude among the elements included in the selected row number of the matrix and is included in each selected column number for each selected column number of the matrix The feature data corresponding to the row numbers including the upper N elements in the order of the size of the elements can be extracted.

The apparatus generates submatrices in which the elements included in each column are sorted in descending order of size for each selected column number of the matrix, and generates submatrices from the first row to the first column of the generated submatrix, Corresponding to the row numbers including the elements extracted from the respective sub-matrices, in each of the column numbers corresponding to the sub-matrices among the column numbers of the matrix, Data can be extracted.

N is a natural number equal to or less than a square root of the number of extracted data and equal to or less than the number of the input data.

The submatrices are n x N matrices when the number of input data is n.

And a computer readable recording medium recording a computer program for executing a feature selection method using a relational matrix according to an embodiment of the present invention.

According to the feature selection method, apparatus, and computer-readable recording medium using the relational matrix according to an embodiment of the present invention, the computation time can be reduced compared to the conventional sequential backward search method.

In addition, it has the effect of selecting features with higher relevance to the association than the Gini Index and MRMR algorithm, and selecting fewer features even when the accuracy is the same.

In addition, since the relational matrix selects characteristics by judging the degree of contribution to the result in the learning process rather than the value of the characteristic, more accurate features can be extracted because similarity between the selected features is not required, Can be used in unsecured machine learning.

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of the specification, illustrate embodiments of the invention and, together with the description, serve to explain the technical features of the invention.
1 is a flowchart briefly illustrating a method of selecting a feature using a relational matrix according to an embodiment of the present invention.
2 is a diagram illustrating an example of a generated matrix in a feature selection method using a relational matrix according to an embodiment of the present invention.
FIG. 3 is a diagram illustrating an example of selecting a most-counted data in a feature selection method using a relational matrix according to an embodiment of the present invention.
4 is a diagram illustrating an example of extracting feature data in a feature selection method using a relational matrix according to an embodiment of the present invention.
5 is a diagram showing experimental data in which features are extracted using a relational matrix according to an embodiment of the present invention.

The terms first and / or second in this specification are used only for the purpose of distinguishing one element from another. That is, the components are not intended to be limited by the terms.

The components, features, and steps referred to in the specification as " comprising " in this specification are intended to mean that there are corresponding components, features, and steps, and do not preclude the presence of one or more other components, features, steps, and the like Is not.

Includes plural forms as long as it is not specified and specified in the singular form herein. That is, the components and the like referred to in this specification may mean the presence or addition of one or more other components or the like.

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 this invention belongs to be.

That is, terms such as those defined in commonly used dictionaries should be construed as meaning consistent with meaning in the context of the related art, and unless otherwise expressly defined herein, interpreted in an ideal or overly formal sense It does not.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Reference will now be made in detail to the preferred embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to the like elements throughout.

1 is a flowchart briefly illustrating a method of selecting a feature using a relational matrix according to an embodiment of the present invention.

Referring to FIG. 1, a feature selection method using a relational matrix according to an embodiment of the present invention includes a step of setting a matrix (S101), a step of counting (S103), a step of selecting data (S105) (Step S107).

A feature selection method using a relational matrix according to an embodiment of the present invention may be performed by a feature selection apparatus using a relational matrix according to another embodiment of the present invention.

The step of generating a matrix (S101) is a step of generating a matrix composed of row numbers and column numbers set for each of the input data. For example, when n pieces of data are input, an n x n matrix composed of a row number and a column number set for each piece of data may be generated, and an element of the matrix may be initialized to zero.

The counting step S103 is a step of counting the number of occurrences of data appearing together with the specific event using the generated matrix whenever a specific event occurs.

That is, among the row numbers of the matrix, row numbers set for data appearing together with a specific event are selected, column numbers set for data appearing together with a specific event among the column numbers of the matrix are selected, Counting the number of occurrences of data while increasing the value of the elements redundantly included in the selected column numbers among the elements included in the numbers by one.

For example, if A, B, and C events occur together when an event X occurs, the elements of the column numbers set for A, B, and C are incremented by 1 in the row numbers set for the A, B, and C events .

(A, B), (B, A), (B, B), (B, C) ) And (C, C) by 1, thereby increasing the value of the element of the matrix set for the events that occur together each time a specific event is generated.

The step of selecting data (step S105) is a step of selecting data in which the number of times of appearance is the greatest among the input data.

Since the matrix after the counting step S103 is symmetric with respect to the diagonal component, the data in which the number of times of appearance is the largest is included in the diagonal component of the matrix.

That is, the data corresponding to the row number of the element having the largest value among the elements included in the diagonal elements of the matrix can be selected as the most counted data.

For example, in the present invention, since values of elements of the diagonal elements such as (A, A), (B, B) and (C, C) also increase, the element having the maximum value among the elements of the matrix is included in the diagonal element do.

The step of extracting feature data (S107) is a step of extracting feature data associated with the data selected in the step of selecting data (S105).

First, column numbers including upper N elements in the order of magnitude among the elements included in the row number of the element having the largest value among the elements of the matrix can be selected.

Here, N may be arbitrarily set to be smaller than the number of rows or columns of the matrix and larger than the square root of the number of feature data to be extracted. That is, the upper limit of the number of feature data extracted according to N can be determined.

For example, if (C, C) among the elements of the matrix is the element having the largest value, the column numbers including the upper two elements in order of magnitude among the elements included in the C row can be selected. If the top two elements of the elements included in row C are included in (C, A) and (C, C), the selected column numbers are A and C.

Next, for each selected column number, the upper N elements may be extracted in descending order among the elements included in each column number.

For example, two nx 1 submatrices are generated in which the elements included in the selected column A and column C are sorted in descending order according to their sizes, and the nx 1 submatrices included in the submatrix from row 1 column 1 row 2 column 1 column Elements can be extracted. Here, n is the number of input data, and in the above example 3 x 1 sub-matrices can be generated.

After extracting the elements for each sub-matrix, the row numbers of the matrix including the extracted elements may be selected in each of the column numbers corresponding to the sub-matrices among the column numbers of the matrix.

For example, if the elements extracted from the sub-matrix corresponding to column A are 3 and 5, rows with elements 3 and 5 in column A of the matrix may be selected. Since the largest element in each column number of the matrix is included in the diagonal of the matrix, row A and row one of the remaining rows can be selected. Similarly, in row C, a row of C and one of the remaining rows may be selected.

Finally, data corresponding to each selected row number can be extracted as the feature data.

For example, if row A and row B are selected in column A and row B and row C are selected in column C, data corresponding to rows A, B, and C may be extracted as feature data.

2 is a diagram illustrating an example of a generated matrix in a feature selection method using a relational matrix according to an embodiment of the present invention.

FIG. 2A is a graph showing a graph 101 for events 1, 2, 3, 4, and 5 predicted to be associated with an A event.

2B is a diagram illustrating a link 103 for events 1, 2, and 4 when events 1, 2, and 4 occur when an A event occurs.

2C shows an example of increasing the values of the elements included in the matrix 105 generated for the 1, 2, 3, 4 and 5 events when the 1, 2 and 4 events occur.

Referring to FIG. 2, the row number and column number of the matrix may correspond to each event. For example, line numbers I1 to I5 correspond to 1 to 5 events, respectively, and column numbers I1 to I5 correspond to 1 to 5 events, respectively.

(I1, I2), (I1, I4), (I2, I1), (I2, I2), (I2, I4) , I1), (I4, I2) and (I4, I4) may be incremented by one.

FIG. 3 is a diagram illustrating an example of selecting a most-counted event in a feature selection method using a relational matrix according to an embodiment of the present invention.

4 is a diagram illustrating an example of extracting feature events in a feature selection method using a relational matrix according to an embodiment of the present invention.

3 to 4, when the value of the elements of the matrix corresponding to the events that occur together is increased every time the event A occurs, the element having the maximum value among the elements of the matrix may be included in the diagonal of the matrix have.

For example, when events 1, 2 and 4 occur together when the event A occurs, the values of the elements included in (I1, I1) as well as (I1, I2) The element having the maximum value among the elements included in the matrix is included in the diagonal element.

In this example, since the element included in (I4, I4) 107 has a value of 13, row I4 is selected and four events corresponding to row I4 can be extracted as the most counted data.

Next, columns including the top N elements in descending order of magnitude of the elements included in row I4 may be selected.

In this example, N is set to three, and columns including the top three elements are selected, and the selected column numbers are I2, I3, and I4.

Finally, row numbers including the top N elements in descending order of the size of the elements 111 included in each selected column number can be selected.

In this example, in each of the columns I2, I3, and I4, row numbers including the top three elements in descending order of the size of the elements included in each column number can be selected.

I2, I4 and I1 rows are selected for the I2 column, I3, I4 and I2 rows for the I3 column are selected, and I4, I2 and I3 rows for the I4 column are selected.

Selected row numbers may be extracted as feature events associated with A events, such as I1, I2, I3, and I4, with 1, 2, 3, and 4 events corresponding to each row number.

5 is a diagram showing experimental data in which features are extracted using a relational matrix according to an embodiment of the present invention.

Dataset Number of Genes Classes Training Data Colon Cancer Database 2000 Normal 22 Cancer 40

Table 1 shows the Colon Cancer Dataset used in (Non-Patent Document 1), Colon Cancer Dataset used in the Gini index, Maximum Relevance and Minimum Redundancy (MRMR), and Relational Matrix according to one embodiment of the present invention. It was used as Dataset for accuracy analysis.

Referring to FIG. 5, a 2000 x 2000 matrix is generated for a total of 2000 Gens, and 75 Gens are extracted using a 20 x 20 relational matrix.

Here, the size of 20 is a number that selects the upper 20 elements in order of size of the elements included in the row number or column number of the matrix, and may be a size that defines the upper limit of the number of feature Gene extracted.

For example, when the value of the element of the Ix row among the diagonal elements of the 2000 x 2000 matrix is the largest, the column numbers including the upper 20 elements in order of magnitude among the elements included in the Ix row can be extracted.

Then, row numbers including the top 20 elements are extracted from among the elements included in each column in the order of the extracted column numbers, and 75 Gens corresponding to the extracted row numbers can be extracted.

In this way, we extracted 14 Gens through a 10 x 10 relational matrix for 75 Gens, and extracted 7 Gens through a 4 x 4 relational matrix for 14 Gens.

Classifier Feature Selection Method Selected Features Count Accuracy (%) Error (%) Bayesian Network Gini Index 10 87.10 12.90 MRMR 10 85.48 14.52 NEWFM Gini Index 10 87.10 12.90 MRMR 10 88.71 11.29 Relational Matrix 7 88.71 11.29

Table 2 shows experimental data on the accuracy of the Colon Cancer Dataset with Gini Index, MRMR, and relational matrix for less than 10 selected features.

As a verification method, a leave-one-out (LOO) cross validation disclosed in (Non-Patent Document 1) was used.

Experimental results show that the combination of the Neural Network with Weighted Fuzzy Membership Functions (NEWFM), MRMR and Relational Matrix has the highest accuracy with an accuracy of 88.71%, NEWFM and Relational Matrix Has selected the smallest number of features by selecting seven features.

As the number of selected features increases, the reliability increases. Even though the accuracy is the same, the combination of NEWFM and relational matrix has the best performance.

Classifier Feature Selection Method Selected Features Count Accuracy (%) Error (%) Bayesian Network
Gini Index 20 88.71 11.29 MRMR 20 83.87 16.13 NEWFM Relational Matrix 14 90.32 9.68

Table 3 shows experimental data on the accuracy of the Colon Cancer Dataset over 10 to 20 selected features selected by Gini Index, MRMR and relational matrix.

As a method of verification, a leave-one-out (LOO) cross validation disclosed in (Non-Patent Document 1) was used as in Table 2.

Experimental results show that the combination of NEWFM and relational matrix has the highest accuracy with accuracy of 90.32% and selects the fewest features by selecting 14 features.

Therefore, the combination of NEWFM and relational matrices shows the best performance when we combine the above results.

Feature Selection Method Index of the selected 10 features (genes) Gini Index 245, 249, 267, 493, 513, 1423, 1671, 1771, 1772 MRMR 249, 493, 513, 625, 765, 1042, 1671, 1771, 1772 Relational Matrix 82, 245, 249, 286, 493, 765, 1772

Table 4 shows the extracted features (genes) according to the combination of NEWFM and Gini Index, MRMR, and relational matrix in Table 2.

249, 493, 765, and 1772 were extracted through three methods. However, 245 was extracted only from Gini Index and relational matrix, and 72 and 286 were extracted from relational matrix only.

Unlike the Gini Index and MRMR, relational matrices do not require the similarity between the selected features because they select the features by judging the extent of contribution to the results in the learning process, not the feature values.

Steps, processes, logic blocks, means, steps, or combinations thereof, in connection with the embodiments described herein may be implemented as electronic hardware (digital designs designed by coding or the like), software Or a combination thereof). Hardware, and / or software may vary depending on design constraints imposed on the user terminal. In addition, the functions described herein may be implemented by hardware, software, firmware, or a combination thereof. When executed by software, the function may be stored on or transmitted via a computer-readable storage medium as one or more instructions or code. Computer-readable storage media generally refers to any available media that can be accessed by a computer.

Although the description herein has been made in some illustrative aspects, various modifications and variations can be made from the categories defined by the following claims, and the technical scope of the invention is defined in the following claims It should be decided by.

101: Graph for 1, 2, 3, 4, and 5 events
103: Link to events 1, 2 and 4
105: Matrix generated for events 1, 2, 3, 4, and 5
107: (I4, I4)
109: Columns containing the top three elements
111: Elements contained in each selected column

Claims (17)

Generating a matrix including a row number and a column number set for each of the input data;
Counting the number of occurrences of data appearing together with the specific event using the generated matrix every time a specific event occurs;
Selecting data in which the number of occurrences is the highest among the input data; And
Extracting feature data associated with the selected data,
Wherein the counting step comprises:
Selecting row numbers set for data appearing together with the specific event among the row numbers of the matrix;
Selecting column numbers set for data appearing together with the specific event among column numbers of the matrix; And
And counting the number of occurrences of the data while increasing the value of the elements included in the selected column numbers among the elements included in the selected row numbers in the matrix.
Feature Selection Method Using Relational Matrix. The method according to claim 1,
Wherein the generating comprises:
Generating a matrix of size nxn if the number of input data is n,
Feature Selection Method Using Relational Matrix. delete The method according to claim 1,
Wherein the step of selecting the data comprises:
Selecting a row number of an element having a largest value among elements included in a diagonal component of the matrix; And
And selecting data corresponding to the selected row number.
Feature Selection Method Using Relational Matrix. 5. The method of claim 4,
Wherein the extracting of the feature data comprises:
Selecting column numbers including upper N elements in order of magnitude among elements included in the selected row number of the matrix; And
Extracting feature data corresponding to the row numbers including the upper N elements in order of magnitude among the elements included in the selected column numbers for each selected column number of the matrix;
Feature Selection Method Using Relational Matrix. 6. The method of claim 5,
Wherein the extracting of the feature data comprises:
Generating sub-matrices in which the elements included in the respective columns are sorted in descending order of size for each selected column number of the matrix;
Extracting elements included in the partial matrixes from the first row and the first column to the N row and the first column of each of the generated sub-matrices; And
Extracting feature data corresponding to row numbers including elements extracted from the sub-matrices in column numbers corresponding to the sub-matrices among column numbers of the matrix,
Feature Selection Method Using Relational Matrix. 6. The method of claim 5,
Wherein N is
Wherein the number of the extracted data is greater than a square root of the number of extracted data,
Feature Selection Method Using Relational Matrix. The method according to claim 6,
The sub-
If the number of the input data is n,
< RTI ID = 0.0 > nx <
Feature Selection Method Using Relational Matrix. Generates a matrix including a row number and a column number set for each of the input data,
Selecting row numbers set for data appearing together with the specific event among the row numbers of the matrix every time a specific event occurs and for detecting data appearing together with the specific event among column numbers of the matrix Counting the number of occurrences of the data while incrementing values of elements redundantly included in the selected column numbers among the elements included in the selected row numbers in the matrix,
Selects the data in which the number of appearance times of the input data is the highest,
And a controller for extracting feature data associated with the selected data,
Feature selection device using relational matrix. 10. The method of claim 9,
Generating a matrix of size nxn when the number of input data is n,
Feature selection device using relational matrix. delete 10. The method of claim 9,
Selecting a row number of an element having a largest value among elements included in the diagonal component of the matrix,
Selecting data corresponding to the selected row number,
Feature selection device using relational matrix. 13. The method of claim 12,
Selecting column numbers including upper N elements in order of magnitude among elements included in the selected row number in the matrix,
Extracting feature data corresponding to the row numbers including the upper N elements in order of magnitude among the elements included in the selected column numbers for each selected column number in the matrix,
Feature selection device using relational matrix. 14. The method of claim 13,
Generating sub-matrices in which the elements included in the respective columns are sorted in descending order of size for each selected column number of the matrix,
Extracts the elements included in the partial matrixes from the first row to the first column of the generated sub-matrix to the N-th row and the first column,
Extracting feature data corresponding to row numbers including the elements extracted from the sub-matrices in each of the column numbers corresponding to the sub-matrices among the column numbers of the matrix,
Feature selection device using relational matrix. 14. The method of claim 13,
Wherein N is
Wherein the number of the extracted data is greater than a square root of the number of extracted data,
Feature selection device using relational matrix. 15. The method of claim 14,
The sub-
If the number of the input data is n,
< RTI ID = 0.0 > nx <
Feature selection device using relational matrix. A computer-readable recording medium recording a computer program for executing the method of claim 1.

Images