JP5667004B2 - Data classification apparatus, method and program - Google Patents

Description

  The present invention relates to a data classification apparatus, method, and program, and more particularly, to a data classification apparatus, method, and program for classifying data using a classification score.

  As a method of simultaneously using the outputs of multiple classification score calculation means when classifying input data, calculate the certainty factor indicating that the class showing the maximum classification score is the class for the input data, and the maximum certainty factor There is a classification method in which a class to be output is displayed as a class to which input data belongs (for example, see Non-Patent Document 1).

Toshio Uchiyama, Katsuto Bessho, Satoshi Uchiyama, Masahiro Oku, "Combination of multiple classifiers using confidence estimation", Japanese Society for Artificial Intelligence, Intelligent Base System Study Group, January 2009.

  However, in the above-described prior art, the technique of Non-Patent Document 1 is that for each classification score calculation means, the probability that the class indicating the first classification score is correct from the classification score of the classification score calculation means (= confirmation) Degree). For this reason, there is a concern that the classification score of other classification score calculation means cannot be reflected in the calculation of the certainty factor, and therefore the prediction accuracy decreases.

  The present invention has been made in view of the above points, and is a data that enables the selection of the classification score calculation means more accurately than the method of predicting the certainty of each classifier output from only the classification score for each classification score calculation means. An object is to provide a classification apparatus, method, and program.

In order to solve the above problems, the present invention (Claim 1) is a data classification device for classifying predetermined input data into classes,
Calculation parameter storage means for storing a relative certainty factor calculation parameter obtained in advance;
A plurality of classification score calculation means having different classification methods or components and features to be calculated;
Classification score storage means for storing the classification score;
A relative confidence calculating means for calculating the relative confidence,
Data class determining means for determining a class or class group to which the input data belongs;
Have
The classification score calculation means includes:
A likelihood score or likelihood from the separation hyperplane that the predetermined input data belongs to each of a plurality of classes, or a classification score that indicates that the larger the value, the higher the probability of belonging to the corresponding class. Means for calculating and storing in the classification score storage means;
The relative certainty factor calculating means includes:
All classification scores output by all the classification score calculation means are acquired from the classification score storage means, and the relative certainty factor calculation parameters stored in the calculation parameter storage means are obtained from all the classification scores. using, there are correct class in the top N (N ≧ 1) corresponding class or classes in each classification score and corresponds to the top N classification score another class score calculating means Including means for calculating relative confidence representing the probability that there is no correct class in the class or class group,
The data class determining means includes
Means for determining a class or class group to which the input data belongs based on the classification score of the classification score calculation means having the highest relative certainty factor

In the present invention (Claim 2), in the relative certainty factor calculating means,
When N class groups are determined, means for calculating using the top N + 1 classification scores of all classification score calculation means is included.

In the present invention (Claim 3), in the relative confidence calculation means,
Use machine learning techniques based on 2-class or multi-class logistic regression (including those using kernels), support vector machines, or other identification models.

The present invention (Claim 4) is a data classification method for classifying predetermined input data into classes,
A plurality of classification score calculation means for calculating classification methods or different constituent elements and features from each other, a likelihood or a likelihood that the predetermined input data belongs to each of a plurality of classes, or a distance from a separation hyperplane There is a classification score calculating step for calculating a classification score indicating that the larger the value is, the higher the probability of belonging to the corresponding class is, and storing the classification score in the classification score storage unit;
Relative confidence calculation means acquires all classification scores output by all the classification score calculation means from the classification score storage means, and stores the relative confidence stored in the calculation parameter storage means from all the classification scores . with confidence factor computing parameters, there are correct class in the top N (N ≧ 1) corresponding class or classes in each classification score, and the higher the classification score another class score calculating means A relative certainty calculating step for calculating a relative certainty representing the probability that there is no correct class among the classes or classes corresponding to N;
The data class determining means performs a data class determining step of determining a class or class group to which the input data belongs based on the classification score of the classification score calculating means having the highest relative certainty factor.

In the present invention (Claim 5), when N class groups are determined in the relative confidence calculation step, calculation is performed using the top N + 1 classification scores of all the classification score calculation means.

Further, the present invention (Claim 6), in the relative confidence calculation step,
Use machine learning techniques based on 2-class or multi-class logistic regression (including those using kernels), support vector machines, or other identification models.

  The present invention (Claim 7) is a data classification program for causing a computer to function as each means of the data classification apparatus according to any one of Claims 1 to 3.

  The present invention uses all the classification score information output by a plurality of classification score calculation means to obtain a relative certainty factor reflecting the classification score of another classification score calculation means, and the classification score calculation means having a high value Is more accurate than predicting the certainty of the output of each classification score calculation means from only the classification score for each classification score calculation means, and enables an accurate classification score calculation means to be selected as a result. A highly accurate classification result can be obtained.

It is a block diagram of the data classification device in one embodiment of this invention. It is a flowchart of a process of the data classification device in one embodiment of this invention.

  Embodiments of the present invention will be described below with reference to the drawings.

  The present invention inputs all the classification scores output by a plurality of classification score calculation means, and predicts the relative certainty of each classification score calculation means as an identification problem.

  FIG. 1 shows a configuration of a data classification apparatus according to an embodiment of the present invention.

  The data classification device 10 shown in the figure includes an input unit 11, a classification score calculation control unit 12, a plurality of classification score calculation means 13, a relative certainty calculation unit 14, a data class determination unit 15, a memory 16, and a classification score storage unit 17. The calculation parameter storage unit 18, the relative certainty storage unit 19, and the class group number storage unit 20, the processing unit storage unit 1 and the keyboard 2 are connected to the input unit 11, and the display 3 is connected to the output unit 15. Has been.

  The processing target storage unit 1 is a database in which processing targets such as documents are stored, and is read by the input unit 11.

  The memory 16 stores the processing target input by the input unit 11.

The calculation parameter storage unit 18 stores a relative certainty factor calculation parameter obtained in advance by a model parameter estimation procedure in a logistic regression model described later.
And step.

  The relative certainty factor storage unit 19 stores the relative certainty factor obtained by the relative certainty factor calculation unit 14.

  The class number storage unit 20 stores class numbers corresponding to the top N classification scores input by the relative certainty factor calculation unit 14.

  The input unit 11 reads the processing target data from the processing target storage unit 1 and stores it in the memory 16, acquires the number of classification score calculation means 13 and the number of classes that are classification destinations input from the keyboard 2, and the classification score It passes to the calculation control unit 12.

  The classification score calculation control unit 12 determines the classification score calculation unit 13 that inputs the feature amount of the processing target data from the memory 16 and calculates the classification score.

  The classification score calculation means 13 calculates a classification score such as a posterior probability or likelihood that the input data belongs to each of a plurality of classes or a distance from the separation hyperplane, and stores the classification score in the classification score storage unit 17. The respective classification score calculation means 13 are different from each other in classification method or constituent elements and features. For example, by modeling the joint probability distribution of data feature vectors and class labels, and directly modeling class posterior probabilities based on Bayesian rules, a generation approach for estimating data class labels and direct model posterior probabilities There is an identification approach.

The relative certainty factor calculation unit 14 uses the relative certainty factor calculation parameter stored in the calculation parameter storage unit 18 from all the classification scores stored in the classification score storage unit 17, and ranks higher for each classification score calculation unit. There are correct classes among the classes or class groups corresponding to N (N ≧ 1), and there are correct classes among the classes or class groups corresponding to the top N classes of other classification score calculation means. Relative certainty representing the probability of not being calculated is calculated and stored in the relative certainty storage unit 19.

  The data class determination unit 15 acquires the relative certainty factor from the relative certainty factor storage unit 19, and determines the class or class group to which the input data belongs based on the classification score of the classification score calculation means that indicates the highest relative certainty factor. ,Output.

  FIG. 2 is a flowchart of the processing of the data classification device according to the embodiment of the present invention, and will be described as a case where the number of class groups to which input data belongs is N (≧ 1).

  Step 101) The input unit 11 reads input data to be processed from the processing target storage unit 1 into the memory 16.

  Step 102) The input unit 11 acquires the number n of classification score calculation means 13 and the number K of classes input from the keyboard 2, and passes them to the classification score calculation control unit 12.

  Step 103) The classification score calculation control unit 12 initializes the number i of the classification score calculation means 13 to 1 (i = 1).

  Step 104) The classification score calculation control unit 12 proceeds to Step 105 if the number i of the classification score calculation means 13 is i ≦ n, otherwise proceeds to Step 110.

  Step 105) The classification score calculation control unit 12 inputs the feature of the input data W stored in the memory 14 to the i-th classification score calculation means 13i, and the input data is class Ck (k = 1,..., K). ) (A posteriori probability, likelihood, distance from the separation hyperplane, etc., the larger the value, the higher the probability of belonging to the corresponding class). As a calculation method, Reference 1 “Osamu Ueda, Kazuaki Saito,“ The Forefront of Stochastic Model Text Model Research for Multi-Topic Text (1) ”, Information Processing Society of Japan, Journal of Information Processing, Vol. 45, No. 2, pp. 184 -190, February 2004 ”may also be used. The calculated classification score is stored in the classification score storage unit 17.

  Step 106) The class group hi corresponding to the top N classification scores of the i-th classification score calculating means 13i is stored in the class group number storage unit 20.

  Step 107) The number i of the classification score calculation means 13i is set to i + 1, and the process returns to Step 104.

  Step 110) Read the classification score that the input data W belongs to each class from the classification score storage unit 17 and set the upper N + 1 of all the classification score calculation means 13 to Lm (m = 1,..., N + 1), and store the calculation parameter. Relative confidence parameter from section 18

を読み込み、次式により第ｉ番目の分類スコア算出手段１３iの相対確信度Piを求め、相対確信度記憶部１９に格納する。

And the relative certainty factor Pi of the i-th classification score calculating means 13i is obtained by the following formula and stored in the relative certainty factor storage unit 19.

式（１）、（２）は、多クラスのロジスティック回帰式である。算出パラメータを求める場合は、文献２「C. M. ビショップ，"Pattern Recognition and Machine Learning (邦訳：パターン認識と機械学習）"，pp. 208-209 Springer 2006．」に記載されている多クラスロジスティック回帰分析において、全ての分類スコア算出手段１３が算出した上位Ｎ＋１個の分類スコアn×(N+1)を説明変数とし、唯一の分類スコア算出手段のみが上位Ｎ個に対応するクラスあるいはクラス群の中に正解があり、他の分類スコア算出手段には無い場合において、正解があった分類スコア算出手段１３iに対応する要素のみが１で、それ以外が０となる目的変数ベクトル（n次元ベクトル）を定義する。目的変数ベクトルが定義できるデータ（説明変数と目的変数）のみを学習データとして最尤推定によりロジスティック回帰パラメータを求め、これを相対確信度を算出するための算出パラメータとしてもよい。

Equations (1) and (2) are multi-class logistic regression equations. When calculating parameters, the multi-class logistic regression analysis described in Reference 2, “CM Bishop,“ Pattern Recognition and Machine Learning ”, pp. 208-209 Springer 2006.” The top N + 1 classification scores n × (N + 1) calculated by all the classification score calculation means 13 are used as explanatory variables, and only one classification score calculation means is included in the class or class group corresponding to the top N. Defines an objective variable vector (n-dimensional vector) where only the element corresponding to the correct classification score calculation means 13i is 1 and the other is 0 when there is a correct answer and there is no other classification score calculation means To do. Logistic regression parameters may be obtained by maximum likelihood estimation using only data (an explanatory variable and an objective variable) that can define an objective variable vector as learning data, and this may be used as a calculation parameter for calculating relative confidence.

  Step 111) The relative certainty factor Pi (i = 1,..., N) of each classification score calculating unit 13 is read from the relative certainty factor storage unit 19, and the classification score calculating unit having the maximum relative certainty factor is selected.

  Step 112) The class group h corresponding to the top N classification scores output by the selected classification score calculating means is read from the class group number storage unit 20 and output as a class group to which the input data belongs.

  The operation of each component of the data classification device can be constructed as a program and installed in a computer used as the data classification device for execution or distributed via a network.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

DESCRIPTION OF SYMBOLS 1 Processing object memory | storage part 2 Keyboard 3 Display 10 Data classification apparatus 11 Input part 12 Classification score calculation control part 13 Classification score calculation means 14 Relative reliability calculation part 15 Data class determination part 16 Memory 17 Classification score memory | storage part 18 Calculation parameter memory | storage part 19 Relative certainty storage unit 20 Class group number storage unit

Claims (7)

A data classification device for classifying predetermined input data into classes,
Calculation parameter storage means for storing a relative certainty factor calculation parameter obtained in advance;
A plurality of classification score calculation means having different classification methods or components and features to be calculated;
Classification score storage means for storing the classification score;
A relative confidence calculating means for calculating the relative confidence,
Data class determining means for determining a class or class group to which the input data belongs;
Have
The classification score calculation means includes:
A likelihood score or likelihood from the separation hyperplane that the predetermined input data belongs to each of a plurality of classes, or a classification score that indicates that the larger the value, the higher the probability of belonging to the corresponding class. Means for calculating and storing in the classification score storage means;
The relative certainty factor calculating means includes:
All classification scores output by all the classification score calculation means are acquired from the classification score storage means, and the relative certainty factor calculation parameters stored in the calculation parameter storage means are obtained from all the classification scores. using, there are correct class in the top N (N ≧ 1) corresponding class or classes in each classification score and corresponds to the top N classification score another class score calculating means Including means for calculating relative confidence representing the probability that there is no correct class in the class or class group,
The data class determining means includes
A data classification device comprising: means for determining a class or class group to which input data belongs based on a classification score of a classification score calculation means having a high relative certainty factor. The relative certainty factor calculating means includes:
2. The data classification apparatus according to claim 1, further comprising means for calculating using the top N + 1 classification scores of all classification score calculation means when determining N class groups. The relative certainty factor calculating means includes:
3. The data classification apparatus according to claim 1, wherein a machine learning method based on two-class or multi-class logistic regression (including those using a kernel), a support vector machine, or another identification model is used. A data classification method for classifying predetermined input data into classes,
A plurality of classification score calculation means for calculating classification methods or different constituent elements and features from each other, a likelihood or a likelihood that the predetermined input data belongs to each of a plurality of classes, or a distance from a separation hyperplane There is a classification score calculating step for calculating a classification score indicating that the larger the value is, the higher the probability of belonging to the corresponding class is, and storing the classification score in the classification score storage unit;
Relative confidence calculation means acquires all classification scores output by all the classification score calculation means from the classification score storage means, and stores the relative confidence stored in the calculation parameter storage means from all the classification scores . with confidence factor computing parameters, there are correct class in the top N (N ≧ 1) corresponding class or classes in each classification score, and the higher the classification score another class score calculating means A relative certainty calculating step for calculating a relative certainty representing the probability that there is no correct class among the classes or classes corresponding to N;
A data class determining step, wherein the data class determining means determines a class or a class group to which the input data belongs based on the classification score of the classification score calculating means with the highest relative certainty;
The data classification method characterized by performing. 5. The data classification method according to claim 4, wherein when N class groups are determined in the relative certainty calculation step, calculation is performed using the top N + 1 classification scores of all classification score calculation means. In the relative certainty factor calculating step,
6. The data classification method according to claim 4, wherein a machine learning method based on two-class or multi-class logistic regression (including those using a kernel), a support vector machine, or another identification model is used. Computer
The data classification program for functioning as each means of the data classification device of any one of Claims 1 thru | or 3.

Images