KR102025280B1 - Method and apparatus for selecting feature in classifying multi-label pattern - Google Patents

Abstract

Disclosed are a method of selecting a feature for classifying multi-labeled patterns, and an apparatus thereof. According to one embodiment of the present invention, the apparatus of selecting a feature for classifying multi-labeled patterns is configured to perform a method of selecting a feature for classifying multi-labeled patterns. The method comprises the following steps: generating an initial feature set having a first feature, as a component, selected according to predetermined restriction conditions from a feature full set including a plurality of features, each of which forms a plurality of patterns capable of being classified by multiple labels; generating a first feature sub set on the basis of the first features of the initial feature set; calculating a relevance evaluation value for each of a plurality of second features not belonging to the first feature sub set in the feature full set and generating a second feature sub set on the basis of the relevance average value; generating a third feature sub set on the basis of the first feature sub set and the second feature sub set; and calculating a suitability value with respect to the third feature sub set and generating a final feature sub set on the basis of the suitability value. According to the present invention, labels of the patterns can be classified by using a specific sub set, and thereby the accuracy of classifying the multi-labeled patterns can be improved.

Description

A feature selection method and apparatus for classifying multiple label patterns {METHOD AND APPARATUS FOR SELECTING FEATURE IN CLASSIFYING MULTI-LABEL PATTERN}

The present invention relates to a feature selection method and apparatus for classifying multiple label patterns, and more particularly, to a feature selection method and apparatus for classifying multiple label patterns that can effectively classify multiple labels in consideration of constraints on a subset of features. It is about.

Recently, many studies have been conducted on multi-label data. Multi-label data is data having one or more labels in one pattern, and has been generated and studied in many fields such as document classification, real-time image classification, genetic information classification, and user emotion classification.

Representative multi-label data is tag information of web document. In order to classify web documents, a web document has tag information and categories are divided based on it. Many documents do not belong to one category but may belong to various categories. For example, an article related to the movie "DaVinci Code", which deals with the issue of religious beliefs, is a document that can belong to both the film category and the religious category.

In this regard, studies have been actively conducted to select features highly correlated with labels in multiple label data (patterns). However, due to the nature of the multiple label problem that requires consideration of several labels in order to calculate the importance of the feature, it is difficult to infer an accurate correlation for the high-dimensional label.

Accordingly, there is a need for developing a technology for improving the classification accuracy of multi-label data.

In this regard, there is a Korean Patent No. 10-1656604 entitled "Method and Device for Selecting a Feature Set Used to Classify Multiple Labels".

SUMMARY OF THE INVENTION The present invention has been made in an effort to provide a method and apparatus for selecting a feature for classifying multiple label patterns for improving the classification accuracy of the multiple label patterns.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the following description.

In a method of selecting a feature for multiple label pattern classification by a multi-label feature selection device according to an embodiment of the present invention for solving the technical problem, a plurality of constituting a plurality of patterns that can be classified into multiple labels Generating an initial feature set comprising the first feature selected according to a predetermined constraint in the entire feature set including the feature, and generating a first feature subset based on the first features of the initial feature set. Generating a relevance evaluation value for each of a plurality of second features not belonging to the first feature subset among the entire feature set, and generating a second feature subset based on the relevance evaluation value; Generating a third feature subset based on the first feature subset and the second feature subset; Calculating a goodness of fit value for the feature subset, and generating m final feature subsets based on the goodness of fit value.

Preferably, the constraints include a maximum number n of features to be included in each feature set or feature subset (n is a natural number), a total feature set or a number of feature subsets m (m is a natural number), and Generating an initial feature set includes generating m initial feature sets each including n or less first features and evaluating each of said initial feature sets.

Preferably, the first feature subset can be generated by applying a genetic operator to the initial feature set.

Advantageously, generating said second feature subset comprises: said first correlation in said first correlation function defining a correlation between said second feature and said first label using a first correlation information measure It may be performed based on a feature correlation function by subtracting a second correlation function defining a correlation between the first feature and the second feature using an information measure.

Preferably, the feature correlation function may be defined by the following equation.

[Equation]

은 제1 상관 관계 함수,

은 제2 상관 관계 함수를 의미할 수 있다. Where l is the label, L is the label set, M is the mutual information measure, and the correlation of the input variables, f _i Is the second feature, f is the first feature,

Is the first correlation function,

May mean a second correlation function.

Preferably, the generating of the third feature subset may further include evaluating the third feature subset using an FFC corresponding to (the number of the second feature subsets). .

Preferably, after generating the final feature subset, generating the second feature subset, generating a third feature subset, and generating the final feature subset until all of the preset FFCs are used. It may further comprise the step of repeating the step of generating.

In the feature selection apparatus for classifying a multi-label pattern according to another embodiment of the present invention for solving the above technical problem, in the entire feature set comprising a plurality of features constituting each of a plurality of patterns that can be classified into multiple labels A feature set generation unit that generates an initial feature set having the selected first feature as a component according to preset constraints, and a first feature that generates a first feature subset based on the first features of the initial feature set A subset generation unit configured to calculate a relevance evaluation value for each of a plurality of second features not belonging to the first feature subset among the set of features, and generate a second feature subset based on the relevance evaluation value A feature subset generator, generating a third feature subset based on the first feature subset and the second feature subset; Calculating a goodness of fit value for the third feature subset generator, the third feature subset, and based on the fitness value comprises a final feature subset generator for generating a subset m of the final characteristics.

According to the present invention, the introduction of a new subset of features that adds promising but unselected features to an entity may improve the performance of population-based search such as GA for constrained multi-label feature selection. Can be.

In addition, since the label of the pattern is classified using a feature subset consisting of only the features selected as the optimal feature, the classification accuracy of the multi-label pattern is improved.

In addition, the amount of calculation and calculation time for calculating the feature evaluation value of the features are reduced, and even when the number of patterns is not large enough, the accuracy of the feature evaluation value calculation result is higher than in the prior art, so that the accuracy of the classification result of the multi-label pattern is higher than in the conventional art. There is an effect to be improved.

Effects of the present invention are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a view for explaining a feature selection method for classifying multiple labels according to an embodiment of the present invention.
2 is a diagram illustrating an apparatus for selecting multiple label features according to an embodiment of the present invention.
3 is a flowchart illustrating a feature selection method for classifying multiple labels according to an embodiment of the present invention.
4 is an algorithm for explaining a feature selection method for classifying multiple labels according to an embodiment of the present invention.
5 is an algorithm for explaining an exploration operator according to an embodiment of the present invention.
6 is a graph showing a result of comparing convergence between GA and the present method according to the number of FFCs used in terms of multi-label accuracy.
7 shows box plots of goodness-of-fit values provided by two descendant sets of a Genbase data set.

As the invention allows for various changes and numerous embodiments, particular embodiments will be illustrated in the drawings and described in detail in the written description. However, this is not intended to limit the present invention to specific embodiments, it should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention. In describing the drawings, similar reference numerals are used for similar elements.

Terms such as first, second, A, and B may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component. The term and / or includes a combination of a plurality of related items or any item of a plurality of related items.

When a component is referred to as being "connected" or "connected" to another component, it may be directly connected to or connected to that other component, but it may be understood that other components may be present in between. Should be. On the other hand, when a component is said to be "directly connected" or "directly connected" to another component, it should be understood that there is no other component in between.

The terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting of the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, the terms "comprise" or "have" are intended to indicate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, and one or more other features. It is to be understood that the present invention does not exclude the possibility of the presence or the addition of numbers, steps, operations, components, components, or a combination thereof.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in the commonly used dictionaries should be construed as having meanings consistent with the meanings in the context of the related art and shall not be construed in ideal or excessively formal meanings unless expressly defined in this application. Do not.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention improves the performance of population-based search such as GA for constrained multi-label feature selection by introducing a new subset of features that adds promising but unselected features to the entity.

1 is a view for explaining a feature selection method for classifying multiple labels according to an embodiment of the present invention.

Referring to FIG. 1, some key issues to consider when introducing a new feature into an evolutionary search based multi-label feature selection process with a constraint are shown.

First, given the full feature set F as shown in (a), we generate chromosomes (feature sets) according to constraints using genetic algorithms as shown in (b). Here, the constraint may include the maximum number n of features to be included in each chromosomes (feature set) and the number m of total chromosomes. Feature set F may have a subset of important features strongly associated with multiple labels, leading to good discrimination when multiple label classifiers are included in the final feature subset. When the random initialization process is completed through the genetic algorithm, important features such as f ₁ may not be selected by any chromosomes. This is because each chromosomes (feature set) handles only a number of features below the constraint n (the number of features to be included in the feature subset). | F | / n chromosomes (feature sets) need to be evaluated with at least all features considered, although all chromosomes are forced to select separate feature sets. This leads to expensive calculation costs. Instead, the present invention can identify promising features with the aid of filters employed without explicit evaluation of candidate feature subsets.

Next, as shown in (c), genetic operators such as crossover and mutation are applied to the population to make new chromosomes. However, because the exchange of ancestor alleles creates new chromosomes, important features that are not selected may not be considered. In other words, a feature that is not selected by an ancestor does not select that feature. The only chance to add an unselected feature to the child generation process is to use a mutation operation. However, this is done by randomly selecting features to achieve convergence and is computationally inefficient because the mutation rate is set to a small value. Therefore, it is necessary to spend a large number of iterations or generations to randomly introduce important features into the population.

Thus, the present invention may apply a search operator to each new offspring as shown in (d) to generate a new subset of features including promising features not considered by the original offspring. During each exploration operation, the dependencies of the unselected features at different labels (l ₁ , l ₂ , ..., l ₈ ) are calculated. After the rank of each feature is calculated (e.g., f ₁ → f ₄₄ → f ₃₂ → f ₃ →...), The most promising feature is selected to generate a new subset of features.

Finally, exploration and genetic operator based feature subsets are integrated into a single population.

2 is a diagram illustrating an apparatus for selecting multiple label features according to an embodiment of the present invention.

2, the multi-label feature selection apparatus 100 according to an embodiment of the present invention may include a feature set generator 110, a first feature subset generator 120, and a second feature subset generator ( 130, a third feature subset generator 140, and a final feature subset generator 150.

The feature set generating unit 110 includes, as a component, a plurality of features constituting each of a plurality of patterns that can be classified into multiple labels as components, m (m is a natural number). Create) initial set of features. In this case, the feature set generation unit 110 generates m feature sets each including n or less first features. Here, n and m may be constraints. Each of the plurality of patterns may be classified into multiple labels, and each of the patterns may include a plurality of features. Here, the pattern may be a document or the like as the object of classification, the label may be a category (genre), and the feature may be a word. In addition, the feature subset is data used to classify each pattern by a specific label. If the pattern is classified using all the features constituting each of the plurality of patterns, the feature subset may be irrelevant or overlapping. Since the use of all of them reduces the performance of the multi-label classification, the present invention improves the multi-label classification performance by classifying the labels of each pattern by using a feature subset composed of high importance features.

The feature set generator 110 may randomly generate a feature set. The generated feature set may be referred to as an initial population.

For example, when a population size is m, the maximum number of features is set to n, and the maximum FFC is set to v, the feature set generator 110 may generate m feature sets including up to n features. Can be.

When the feature set is generated according to the constraints as described above, the feature set generator 110 evaluates each feature set. In this case, the feature set generator 110 may evaluate each feature set by using a preset FFC.

The first feature subset generator 120 generates a first feature subset based on the first features of the feature set. In this case, the first feature subset generator 120 may generate a first feature subset by applying a genetic operator to each feature set.

The second feature subset generation unit 130 calculates a relevance evaluation value for each of the plurality of second features not belonging to the first feature subset among the feature set, and based on the relevance evaluation value, the second feature subset Create Since the first subset of features selects a small number of first features within n and most of the features remain unselected, a search operator is required to search for the features that are not searched. Accordingly, the second feature subset generator 130 may generate a second feature subset by applying an exploration operator to the first feature subset.

Specifically, the second feature subset generator 130 maximizes the objective function for each first feature subset generated by the genetic operator, and the size of the feature subset becomes | S _c | Recursively select relevant features not selected by offspring c until. Where | S _c | is the subset size of c. Thus, the exploration operator does not need additional parameters to determine the number of features to be selected. To perform the exploration operation, we use an effective filter method called scalable criterion for large label (SCLS) as the objective function Q (f ⁺ , L). Where L is the set of labels. The selection of the i th feature in the feature set {F | {S _c ∪ Z}} is performed by identifying the feature f _i that maximizes the relevance evaluation value. Here, Z may be a feature subset having the i−1 th feature when the i th feature is selected.

Therefore, the second feature subset generation unit 130 performs a relevance evaluation using Equation 1 below.

[Equation 1]

는 L에 대한 f_i 의 연관성(dependency),

는 Z의 선택된 특징에 대한 f_i 의 연관성(dependency), f_i 는 i번째 특징을 나타낸다. here,

Is the dependence of f _i on L,

Denotes the dependency of f _i on the selected feature of Z, f _i denotes the i-th feature.

Equation 1 may be reconfigured as in Equation 2 below.

[Equation 2]

으로, 변수 x와 y사이의 상호 정보이고,

로 확률 함수 p(x), p(y) 및 p(x, y)의 결합 엔트로피(joint entropy)일 수 있다. 따라서, 수학식 2로부터 D(f₂)를 아래 수학식 3과 같이 표현될 수 있다. Where l is a label, L is a label set, M (x; y) is a mutual information measure, f _i May be the i th feature. And,

, Mutual information between the variables x and y,

May be a joint entropy of the probability functions p (x), p (y) and p (x, y). Therefore, D (f ₂ ) from Equation 2 may be expressed as Equation 3 below.

 [Equation 3]

는 아래 수학식 4와 같이 표현될 수 있다. Also, from equation (2)

May be expressed as Equation 4 below.

[Equation 4]

In order to consider the adaptability to scaling of D (f ₂ ) and to calculate R (f ₂ ) while avoiding iterative calculations by f∈S and l∈L, R (f ₂ ) is It can be expressed as 5.

[Equation 5]

Here, when 0 ≦ α ≦ 1 to be estimated, a decrease in relevance for f ₂ is determined based on D (f ₂ ) while avoiding repeated calculations for each label. Therefore, α can be approximated as Equation 6 below.

[Equation 6]

In conclusion, the relevance assessment for f ₂ may be calculated using Equation 7 described below.

[Equation 7]

Equation 7 shows how relevance evaluation is performed when i = 2. Considering the previously selected features in Z, the final relevance assessment can be expressed by Equation 8 below.

[Equation 8]

은 제1 상관 관계 함수,

은 제2 상관 관계 함수를 의미할 수 있다. Where l is the label, L is the label set, M is the mutual information measure, and the correlation of the input variables, f _i Is the second feature, f is the first feature,

Is the first correlation function,

May mean a second correlation function.

Equation 8 may be an objective function for calculating a relevance evaluation value for each of a plurality of second features not belonging to the first feature subset, and selecting a feature included in each feature subset based on the relevance value.

Accordingly, the second feature subset generation unit 130 uses the first mutual information measure in the first correlation function to define a correlation between the second feature and the first label using the first mutual information measure. A second subset of features may be generated based on a feature correlation function subtracting a second correlation function defining a correlation between the feature and the second feature. In this case, the feature correlation function may be Equation 8. The size of the second feature subset may be the same as the size of the first feature subset.

The third feature subset generator 140 generates a third feature subset based on the first feature subset and the second feature subset. In this case, the third feature subset generation unit 140 may evaluate the third feature subset by using the FFC corresponding to (number of second to first feature subsets).

The final feature subset generator 150 calculates a goodness of fit value for the third feature subset, and generates m final feature subsets based on the goodness of fit value. That is, the final feature subset generator 150 adds the third feature subset to the initial feature set, and selects the m final feature subsets in order of high evaluation value.

3 is a flowchart illustrating a feature selection method for classifying multiple labels according to an embodiment of the present invention, and FIG. 4 is a view for explaining a feature selection method for classifying multiple labels according to an embodiment of the present invention. 5 is an algorithm for explaining an exploration operator according to an embodiment of the present invention.

Referring to FIG. 3, the apparatus generates m initial feature sets P (t) having m (m is a natural number) as components of the first feature selected from the full feature set (S310), and the initial feature set P (t). ) Is evaluated (S320). That is, the device generates m initial feature set P (t) through random allocation of up to n binary bits of the selected feature from the full feature set. Here, n means the maximum number of features allowed in the feature subset Sc. The device then uses the goodness-of-fit function to evaluate the initial feature set P (t). In this case, the device uses multiple label classification errors as a goodness-of-fit function for the initial feature set P (t). Since m initial feature sets P (t) must be evaluated to obtain a goodness of fit value, n goodness-of-fit calls (FFCs) are used.

When the initialization process is completed, the device generates a first feature subset G (t) by using a genetic operator (S330). At this point, the device generates a first feature subset G (t) that controls the number of selected features using a crossover operator, mutation operator, and the like.

On the other hand, since the first feature subset selects fewer than n features, most features remain unselected. Thus, an exploration operator is needed to explore the unexplored feature.

Accordingly, when step S330 is performed, the apparatus generates a second feature subset E (t) by applying an exploration operator to the first feature subset G (t) (S340). At this time, the device maximizes the objective function for each child produced by the genetic operator and repeatedly selects relevant features not selected by the offspring until the size of the feature subset is | S _c |. Where | S _c | may be a subset size. The size of the second feature subset E (t) may be the same as the size of the first feature subset G (t) to balance the genetic operator and the exploration operator. This may set the size of E (t) equal to the value of G (t) because the second feature subset E (t) must be evaluated to determine its suitability. The algorithm for the search operator may be the same as in FIG. 5. Referring to FIG. 5, the objective function is maximized for each first feature subset generated by the genetic operator, and the offspring _c is turned off until the size of the feature subset becomes | S _c |. Relevant features not selected by the selection are repeatedly selected. Where | S _c | is the subset size of c. Thus, the exploration operator does not need additional parameters to determine the number of features to be selected. To perform the exploration operation, we use an effective filter method called scalable criterion for large label (SCLS) as the objective function Q (f ⁺ , L). Where L is the set of labels. The selection of the i th feature in the feature set {F | {S _c ∪ Z}} is performed by identifying f _i that maximizes the relevance evaluation value. Here, Z may be a feature subset having the i−1 th feature when the i th feature is selected.

When step S340 is performed, the device combines the first feature subset and the second feature subset to generate a third feature subset N (t) (S350) and evaluate the third feature subset (S360). That is, the device generates a third feature subset N (t) of the t th population and evaluates the third feature subset using a certain number of FFCs. In this case, the device may use an FFC of 2 · | G (t) | in one generation.

When step S360 is performed, N (t) is added to P (t), and m feature subsets having a high fitness value are selected (S370).

From step S330 to step S370 is repeated until all the allowed FFCs are used (S380). Through the above process, an optimal subset of features may be selected.

Hereinafter, the effects of the present invention will be described through experimental results.

The present invention tested 20 different data sets in various areas. The Birds data set is audio data that contains examples of several bird calls. The emotional data set is music data classified into six emotional clusters. Enron, language log (LLog), and Slashdot data sets were created in text mining applications, where each feature corresponds to the appearance of a word, and each label represents a relevance to a specific subject in each text pattern. Genbase and Yeast data sets are derived from biological domains and contain information about the function of genes and proteins. Mediamill data set is the video data of the auto-sensing system. Medical data sets were sampled from large data (coopers) of suicide letters obtained from natural language processing of clinical free text.

Scene datasets are associated with semantic indexes of still scenes, where each scene can contain multiple objects. The TMC2007 data set contains safety reports for complex spatial systems. The remaining nine datasets are from the Yahoo dataset collection. Unsupervised dimensionality reduction of text data sets was performed, including the TMC2007 and Yahoo collections, which consist of more than 10,000 features. In particular, the top 2% and 5% of the document-highest features were maintained for the TMC2007 and Yahoo data sets, respectively. In the text mining domain, existing studies report that classification performance is not significantly affected by maintaining 1% of features based on document frequency.

Table 1 shows standard statistics for multiple label data sets used in the experiment.

TABLE 1

Table 1 shows the number of patterns in the data set | W |, feature count | F |, feature type and number of labels | L | If the feature type is numeric, the feature is discretized using the supervised discretization method of the multi-label naive bayes classifier (MLNB). In particular, each observed numerical value is assigned to one of several bins that are automatically determined using the discretization method.

Label cardinality card (label cardinality Card) represents the average number of labels in each instance. Label density Den is the label cardinality for the total number of labels. The number of unique label sets Distinct represents the number of unique label subsets in L. Domain represents the application from which each data set is extracted.

The average size of the selected feature set for this method and the existing multiple label feature selection method (limited genetic operator (RGA), NSGA-II, MPSOFS) was measured. Depending on which method you choose, you can select up to 10 features. In particular, it provides detailed parameter settings that support good reproducibility as follows.

RGA generates m = 20 initial solutions by randomly selecting smaller than n = 10 features for each chromosome. Each solution of the initial population P (t) with t = 0 is evaluated using multiple label classifiers. The RGA then uses a genetic operator to produce an offspring set N (t). To apply the intersection operator, two solutions are randomly selected and matched in P (t). One solution of P (t) is then randomly selected and mutated. In the present invention, a limited crossover and a restriction mutation operator with both crossover and mutation rates set to 1.0 was used. Thus, for each iteration, GA creates three new solutions containing N (t). Each newly created solution is evaluated using multiple label classifiers. To create P (t + 1), N (t) is added to P (t), and 20 solutions with higher goodness-of-fit values are selected. This procedure is repeated until the RGA uses 100 FFCs.

NSGA-II is the same number that RGA generates, randomly generating m = 20 initial solutions. The maximum number of features allowed is set to | F |. This is because NSGA-II naturally minimizes the number of features selected. Each solution of P (t) is evaluated using the number of multiple label classifiers and features employed. NSGA-II then | N (t) | N (t) with = 3 is the same setting for RGA | N (t) | Produces N (t) with = 3. To generate P (t + 1), N (t) is added to P (t) and the superiority of each solution is determined by the nondominated sorting method. {P (t) ∪ N (t)} Once the superiority is determined among the solutions, the top 20 solutions are selected to form P (t + 1). This procedure is repeated until NSGA-II uses 100 FFCs.

MPSOFS is the same number that RGA generates, randomly generating 20 initial solutions. Each solution of P (t) is evaluated using the number of multiple label classifiers and features used, and ranked using an unclassified sorting method. MPSOFS then preserves P (t) 's best solution, a globally optimal solution. It also preserves the best solution experienced by each chromosome. This is called the individual best solution, so there are 20 individual best solutions. MPSOFS then updates the expression of each chromosome based on its global optimal solution and its own individual best solution. In this case, we use a velocity with an inertial weight of 0.7298 and two acceleration coefficients of 1.4962. After all the chromosomes of ?? (??) are deformed, they are evaluated and considered to be P (t + 1). This procedure is repeated until MPSOFS uses 100 FFCs.

On the other hand, performance can be improved due to different parameter settings, but for fair comparison, the size of population m is set to 20 and the number of used FFCs v is set to 100 for all methods. To assess the quality of feature subsets obtained by each method, output predicted subsets of labels based on the unique characteristics of a given dataset without going through complex parameter tuning that can affect the performance of the final multilabel classification. For this reason, an MLNB classifier was used. For fairness, the hold out cross validation method was used for each experiment. Eighty percent of the samples in a given data set were randomly selected as a training set for multilabel feature selection and classifier training, and the remaining 20% was used as a test set to achieve multilabel classification performance. For both RGA and the proposed method, the population size was set to 20 and the maximum allowable FFC number was set to 100. Each experiment was repeated 10 times, and the mean value was used to indicate the classification performance of each feature selection method.

Four evaluation metrics were used: Hamming Loss, Multiple Label Accuracy, Ranking Loss, and Normalized Coverage. T = {(T _i , λ _i ) | Suppose 1 ≤ i ≤ | T |} is given as a test set. Where λ _i ⊆ L is a valid label subset. For a given test sample T _i , a classifier, such as MLNB _, should output a set of confidence values of 0 ≤ ψ _{i, l} ≤ 1 for each label l ∈ L. If the confidence value ψ _{i, l} is greater than a predefined threshold value such as 0.5, then that label l is included in the predicted label subset Y _i . Ground truth λ _i , confidence ψ _{i, l} and predicted label subset Y _i Based on the multiple label classification performance can be measured on each evaluation scale.

Multiple label accuracy is calculated using Equation 9 below.

[Equation 9]

Hamming loss is calculated using Equation 10 described below.

[Equation 10]

Where T is a given test set, λ represents the correct label subset, and Δ represents the symmetric difference between the two sets.

The ranking loss is calculated using Equation 11 below.

[Equation 11]

는

의 상보 집합(complementary set )이다. 랭킹 손실은 모든 관련성이 있고 관련성이 없는 레이블 쌍에 대해

인 (a, b) 쌍의 평균 분율(average fraction )을 측정한다. here

Is

Is a complementary set of. Ranking loss is for all relevant and unrelated label pairs

Measure the average fraction of the phosphorus (a, b) pairs.

Finally, the normalized coverage is calculated using Equation 11 below.

[Equation 12]

Here, the rank (·) returns the rank of the corresponding label l ∈ λ _i in a non-incremental order according to ψ _{i, l} . Thus, the standardized range measures the number of labels that must be expressed as positive so that all relevant labels are positive. Higher multi-label accuracy, lower hamming loss, ranking loss, and normalized coverage values indicate better classification performance.

In addition, to verify the performance of the method, Wilcoxon signed-rank test is performed to confirm the superiority of the method. Let d _{i be} the performance difference between the two methods for the i th dataset. The difference is ranked according to the absolute value, with the smallest d _i assigned as the first rank. If a tie occurs, an average rank is assigned. R ⁺ may be a sum of ranks for a data set in which a compared method defined as in Equation 13 below exceeds the present method.

[Equation 13]

Let R ^{− be} the sum of the ranks for the data sets that outperform the compared methods. Then, when the confidence level is α = 0.05 and N = 20 based on the Wilcoxon test threshold, the difference in the compared methods is important if min (R ⁺ , R ⁻ ) is less than or equal to 8. In this case, the null hypothesis for the same performance is rejected.

Experimental results through the experimental settings as described above are as follows.

First, comparison results will be described.

Table 2 shows the results for the mean size and standard deviation of the selected subset of features in this method and the existing multiple label feature selection method when the evaluation scale is multi-label accuracy.

TABLE 2

In Table 2, the X symbol represents a method that does not meet the given constraints for that data set. The method and RGA chose less than 10 features for all data sets. The NSGA-II and MPSOFS methods did not select less than 10 features for all data sets other than NSGA-II's Mediamill data set, although they had an objective function to minimize feature subset size. Since NSGA-II and MPSOFS did not select less than 10 features in most data sets, the performance of this method was compared with the RGA performance of subsequent experiments. N means setting a value greater than 10, such as 30 or 50. The experimental results in Table 2 show that NSGA-II or MPSOFS does not meet the given constraints because it outputs the final feature subset. The final feature subset then consists of tens or hundreds of features in most experiments.

Tables 3 and 4 contain the experimental results for this method and the RGAs for 20 multiple label data sets and present the average performance for the corresponding standard deviation and holdout cross-validation. Table 3 contains performance results for multi-label accuracy and hamming loss, and Table 4 contains performance results for ranking loss and normalized coverage.

TABLE 3

TABLE 4

In Table 3 and Table 4, the best performance between the two methods is shown in bold and √. Finally, Table 5 shows the results of the Wilcoxon signed-rank test for the present method for RGA on a Genbase data set with a significance threshold of α = 0.05.

TABLE 5

In Table 5, for each rating scale, the winner of each comparison is shown in bold font and the corresponding sum of the R ⁺ and p values that perform well in the total ranking is shown in parentheses. Similar trends can be observed with other multilabel datasets.

As shown in Tables 3 and 4, this method outperforms RGA for most multilabel datasets. Specifically, it can be seen that the method achieves the best performance with 90% multiple label accuracy, 95% of hamming loss, 95% of ranking loss, and 100% in normalized coverage of the data set. Thus, it can be seen that the method significantly outperforms the RGA for all rating measures. This is evident from the experimental results presented in Table 5, which clearly demonstrates that the method is statistically superior to RGA.

Next, the experimental results according to the present invention will be analyzed.

6 is a graph showing a result of comparing convergence between GA and the present method according to the number of FFCs (u) used in terms of multiple label accuracy. In Figure 6, the axis of abscissas is u, and the axis of ordinates is multilabel performance. Convergence can be different for each experiment due to the probabilistic nature of the population-based search method. Therefore, we set the same initialized population in both algorithms and averaged the multi-label accuracy performance of the upper elites in the population after 10 experiments. . 6 shows that the multi-label accuracy performance is monotonically improved to u. The initialization step uses 20 FFCs, and both methods gradually improve multi-label accuracy because the same initialized population is randomly generated. However, the experimental results show that since the exploration operator is applied to the population after initialization, the multi-label accuracy value of the method improves dramatically when u is greater than 20. 6 thus shows that the method can efficiently locate a good subset of features from unselected features.

The goal of the exploration operator is to introduce new and promising features that can effectively improve the performance of multiple label classification. To verify the validity of the exploration operator, an additional experiment was performed to compare the goodness-of-fit values of the descendant set generated by the exploration operator with the goodness-of-fit values of the random operator. Specifically, 50 chromosomes, G, were selected that selected 10 or fewer features as the same initialization procedure of RGA, and an exploration operator was applied to each chromosome of G of the first set of progeny to generate 50 new chromosomes. Thereafter, for comparison, a new feature for each chromosome of G is randomly selected and introduced to generate a second set of offspring. Finally, the goodness-of-fit values of the first and second progeny sets were measured using four performance measures.

7 shows box plots of goodness-of-fit values provided by two descendant sets of a Genbase data set. Experimental results show that the search capability of the exploration operator is much higher than that of random search, because the goodness-of-fit value of the first set of descendants (suggestions) is much higher than the goodness-of-fit value of the second set of children (any) in terms of all measurements. You can see that it is high.

In conclusion, the present invention utilizes an efficient evolutionary search based feature selection method with a constraint for multiple label classification. Because feature subsets select fewer features within the maximum allowable number of features, and most features are not selected in the constraint constraint problem, new exploration is needed to find relevant features in the unselected feature subsets. Operator. Experiments with 20 real data sets showed that the exploration operator successfully improved the retrieval power of gene retrieval, thereby improving multilabel classification. We also showed that the method can successfully retrieve feature subsets, which did not violate the budget constraint. Statistical tests show that the method outperforms the four performance measures compared to the existing method.

On the other hand, embodiments of the present invention can be implemented in the form of program instructions that can be executed by various computer means may be recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the media may be those specially designed and constructed for the purposes of the present invention, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Examples of program instructions such as magneto-optical, ROM, RAM, flash memory, etc. may be executed by a computer using an interpreter as well as machine code such as produced by a compiler. Contains high-level language codes. The hardware device described above may be configured to operate as one or more software modules to perform the operations of one embodiment of the present invention, and vice versa.

So far I looked at the center of the preferred embodiment for the present invention. Those skilled in the art will appreciate that the present invention can be implemented in a modified form without departing from the essential features of the present invention. Therefore, the disclosed embodiments should be considered in descriptive sense only and not for purposes of limitation. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the scope will be construed as being included in the present invention.

100: multi-label feature selection device
110: feature set generation unit
120: first feature subset generation unit
130: second feature subset generator
140: third feature subset generator
150: final feature subset generation unit

Claims (8)

In the multi-label feature selection device to select a feature for the classification of a multi-label pattern,
The feature set generation unit generates an initial feature set including, as a component, a first feature selected according to a preset constraint from a whole feature set including a plurality of features constituting a plurality of patterns that can be classified into multiple labels. step;
Generating, by a first feature subset generation unit, a first feature subset based on the first features of the initial feature set;
The second feature subset generation unit calculates a relevance evaluation value for each of a plurality of second features not belonging to the first feature subset among the feature set, and generates a second feature subset based on the relevance evaluation value. Generating;
Generating, by a third feature subset generation unit, a third feature subset based on the first feature subset and the second feature subset; And
A final feature subset generation unit calculating a goodness of fit value for the third feature subset, and generating a final feature subset based on the goodness of fit value,
The constraint includes a maximum number n of features to be included in each feature set or feature subset (n is a natural number), the total number of features or a total number of feature subsets m (m is a natural number),
Generating the initial feature set,
Generating m initial feature sets each comprising n or less first features; And
And evaluating each of the initial feature sets. delete The method of claim 1,
The first feature subset is
And generating by applying a genetic operator to the initial feature set. In the multi-label feature selection device to select a feature for the classification of a multi-label pattern,
The feature set generation unit generates an initial feature set including, as a component, a first feature selected according to a preset constraint from a whole feature set including a plurality of features constituting a plurality of patterns that can be classified into multiple labels. step;
Generating, by a first feature subset generation unit, a first feature subset based on the first features of the initial feature set;
The second feature subset generation unit calculates a relevance evaluation value for each of a plurality of second features not belonging to the first feature subset among the feature set, and generates a second feature subset based on the relevance evaluation value. Generating;
Generating, by a third feature subset generation unit, a third feature subset based on the first feature subset and the second feature subset; And
A final feature subset generation unit calculating a goodness of fit value for the third feature subset, and generating a final feature subset based on the goodness of fit value,
Generating the second feature subset comprises:
In the first correlation function defining a correlation between the second feature and the first label using a first mutual information measure, the correlation between the first feature and the second feature is determined using the first mutual information measure. A method of selecting a feature for multiple label pattern classification, characterized in that performed based on a feature correlation function by subtracting a second correlation function to be defined. The method of claim 4, wherein
The feature correlation function is defined by the following equation.
[Equation]

Where l is the label, L is the label set, M is the mutual information measure, and the correlation of the input variables, f _i Is the second feature, f is the first feature,

Is the first correlation function,

Means the second correlation function. The method of claim 1,
Generating the third subset of features may include:
And evaluating the third feature subset using the FFC corresponding to (number of second feature subsets). The method of claim 1,
After generating the final subset of features,
And repeating the step of generating the second feature subset, generating the third feature subset, and generating the final feature subset until all of the preset FFCs have been used. Feature selection method for pattern classification. A feature set generation unit for generating an initial feature set including the first feature selected according to a predetermined constraint in the entire feature set including a plurality of features constituting a plurality of patterns that can be classified into multiple labels;
A first feature subset generator configured to generate a first feature subset based on the first features of the initial feature set;
A second feature subset generation for calculating a relevance evaluation value for each of a plurality of second features not belonging to the first feature subset among the set of features, and generating a second feature subset based on the relevance evaluation value part;
A third feature subset generator configured to generate a third feature subset based on the first feature subset and the second feature subset; And
A final feature subset generator configured to calculate a goodness of fit value for the third feature subset, and generate a final feature subset based on the goodness of fit value,
The constraint includes a maximum number n of features to be included in each feature set or feature subset (n is a natural number), the total number of features or a total number of feature subsets m (m is a natural number),
The feature set generation unit generates m initial feature sets each including n or less first features, and evaluates each of the initial feature sets.

Images