JP5353443B2 - Data classifier creation device, data classifier, data classifier creation method, data classification method, data classifier creation program, data classification program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To allow a plurality of evaluators or a plurality of decision criteria to be used for one piece of input data. <P>SOLUTION: In a data classifier creation device 1, first a 1:1 data creation part 20 creates a 1:1 data set being a set of 1:1 data for classifier creation on the basis of data stored in an instance data storage part 11 and stores the 1:1 data set in a 1:1 data storage part 12. A classifier creation part 30 generates a plurality of classifiers on the basis of the 1:1 data set stored in the 1:1 data storage part 12 and stores them in a classifier storage part 13. A classifier output part 40 reads a plurality of classifiers from the classifier storage part 13 and outputs the plurality of classifiers via an external connection part 100. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to a digital content classification technique and a search technique, and relates to a data classifier creation apparatus, a data classifier creation method, a data classifier creation program, and a creation program for creating a classifier for classifying digital contents into classes. The present invention relates to a data classification device, a data classification method, and a data classification program for classifying digital contents based on the classified classifier.

  In recent years, a large amount of digital content has been widely stored and utilized in a large-capacity storage medium, a computer, etc., against the background of the development of compression technology related to digital content and the widespread use of large-capacity storage media. As the number of stored digital contents increases in this way, there is an increasing demand for a technique for searching for desired digital contents. As one of the search methods, there is a method in which digital contents are classified into categories in advance, and a user acquires desired digital contents using the category information. As a method of classifying digital contents into categories, there is a method of using a classifier that performs category classification by an arithmetic process from an input signal, in addition to a method of direct classification by the user himself / herself. Furthermore, Patent Literature 1 and Patent Literature 2 disclose techniques for improving classification performance by combining a plurality of classifiers.

Patent Document 1 describes a technique in which outputs of a plurality of rules / discriminant trees generated by a concept learning unit are input to a majority decision unit, and the output of the majority decision unit is used as a final classification result.
In Patent Document 2, learning is performed such that when a plurality of decision trees are constructed based on learning data and the decision trees are linearly combined by the weighted majority method, the error probability estimated from the empirical error probability is reduced. A system for creating a classifier is described.

Japanese Patent Application Laid-Open No. 7-064793 JP 2001-195379 A

  By the way, as described in the above prior art, when creating a classifier such as a decision tree, a learning data set that stores a plurality of sets of digital content and a desirable classification category corresponding thereto is prepared in advance. There is a need. When classifying digital content, it is often necessary to perform so-called subjective classification in which correct classification results differ depending on the evaluator. In particular, when performing classification related to human sensitivity and preference, it is obvious that this tendency is remarkable, and the correct classification result varies depending on the evaluator. For example, when giving a correct answer about the impression of a certain digital content, one evaluator said that the category “beautiful” was correct, but another evaluator said that the category “cute” was correct. Is the case. Since this occurs because there are individual differences in the evaluation criteria of the evaluator, it cannot be said that both categories are wrong.

  However, in the above prior art, when a training data set is created by multiple evaluators and there are multiple “correct answers” for the same input data, a classifier is created using which learning data. Whether or not to do so was not fully considered.

  Therefore, the present invention provides a data classifier creating apparatus, a data classifier creating method, a data classifier creating program, and a single input data that can handle a plurality of evaluators or a plurality of judgment criteria for one input data. An object of the present invention is to provide a data classification device, a data classification method, and a data classification program capable of classifying digital contents by handling a plurality of evaluators or a plurality of judgment criteria.

In order to solve the above-described problem, the data classifier creating apparatus according to the present invention is case data in which input data corresponds to correct data indicating a correct classification of the input data, and a plurality of the same input data. A case data storage unit for storing a set of case data including cases corresponding to a plurality of correct answer data given by the evaluator, and reading the case data set from the case data storage unit, one input data is 1 While referring to the one-to-one data creation unit that creates a plurality of one-to-one data sets corresponding to one correct data for each of the plurality of evaluators and the created plurality of one-to-one data sets A data classifier creating apparatus including a classifier creating unit that creates a plurality of classifiers by controlling to create one classifier from a set of one-to-one data.
Further , the data classification device of the next invention is input to each of the classifier storage unit storing the plurality of classifiers created by the data classifier generation device and the plurality of classifiers stored in the classifier storage unit The data classification device includes a classification execution unit that inputs data and obtains a plurality of classification results, and a classification result integration unit that outputs a final class using the plurality of classification results .
The data classifier creating a next invention, the input data, a case data and correct answer data corresponding indicating the classification of the input data correct, by a plurality of evaluators for the same input data A set of case data is read from a set of case data including cases corresponding to a plurality of given correct answer data, and a set of one-to-one data in which one input data corresponds to one correct answer data is evaluated. Creating a plurality of numbers for each person and controlling to create one classifier from one set of one-to-one data while referring to the plurality of sets of one-to-one data created, A method of creating a plurality of classifiers.
In the data classification method of the next invention , input data is input to each of the plurality of classifiers stored in the step of storing the plurality of classifiers created by the data classifier creation method and the classifier storage step. A data classification method comprising: a step of inputting and obtaining a plurality of classification results; and a step of outputting a final class using the plurality of classification results .
The data classifier creation program of the next invention is case data in which input data corresponds to correct data indicating a correct answer classification of the input data, and a plurality of evaluators performs the same input data. A set of case data is read out from a set of case data including cases corresponding to a plurality of given correct answer data, and a set of one-to-one data in which one input data corresponds to one correct data is the plurality of evaluators. Creating a plurality of numbers for each, and controlling to create one classifier from one set of one-to-one data while referring to the plurality of sets of one-to-one data created A data classifier creating program for causing a computer to execute the step of creating a classifier of
The data classifying program of the next invention is a method of storing a plurality of classifiers created by the data classifier creating program, and input data to each of the plurality of classifiers stored in the classifier storing step. A data classification program that causes a computer to execute a step of inputting and obtaining a plurality of classification results and a step of outputting a final class using the plurality of classification results .

According to the present invention, it is possible to handle a plurality of evaluators or a plurality of judgment criteria for one input data, and classify digital contents by handling a plurality of evaluators or a plurality of judgment criteria for one input data. can do.
As a result, even if a classifier is constructed with a data set in which a plurality of correct answer data are assigned to one input data, a classifier with high classification accuracy can be created.
Moreover, even if it is correct answer data created by a plurality of evaluators for one input data, it is possible to create a classifier with high accuracy. A classifier can be created.
Also, if multiple classifiers are constructed from learning data consisting of multiple evaluators, and the classification results obtained from the multiple classifiers are integrated to obtain the final classification results, the accuracy will be higher than before. Good classification results can be obtained.

It is a block diagram which shows the structure of the data classifier production apparatus 1 in Embodiment 1,2. 6 is a flowchart showing a flow of processing of the data classifier creating apparatus 1 in the first and second embodiments. It is a figure which shows an example of the case data storage part in Embodiment 1-4. 6 is a flowchart showing a flow of processing of a one-to-one data creation unit in the first and second embodiments. It is a figure which shows an example of the one-to-one data storage part in Embodiment 1,2. It is a flowchart which shows the flow of the shuffle process in Embodiment 1-4. It is a flowchart which shows the flow of a process of the classifier preparation part in Embodiment 1-4. It is a figure which shows an example of the classifier pair in Embodiment 1-4. It is a figure which shows an example of the classifier storage part in Embodiment 1,2. It is a figure which shows an example of the gene individual in Embodiment 1-4. It is a flowchart which shows the flow of the process which produces the classifier by applying the genetic algorithm in Embodiment 1-4. It is a flowchart which shows the flow of the generation change process in Embodiment 1-4. It is a figure which shows an example of the one point crossing in Embodiment 1-4. 6 is a block diagram showing a configuration of a data classification device 2 in Embodiment 2. FIG. 10 is a flowchart illustrating a processing flow of the data classification device 2 according to the second embodiment. 6 is a block diagram showing a configuration of a data classifier creation device 3 in Embodiments 3 and 4. FIG. 10 is a flowchart showing a flow of processing of a one-to-one data creation unit in the third and fourth embodiments. FIG. 10 is a diagram illustrating an example of a one-to-one data storage unit in the third and fourth embodiments. It is a figure which shows an example of the classifier storage part in Embodiment 3, 4. FIG. FIG. 10 is a block diagram showing a configuration of a data classification device 4 in a fourth embodiment. 14 is a flowchart illustrating a processing flow of the data classification device 4 according to the fourth embodiment.

First, the characteristics of the present invention are shown as follows.
[1] A case set with multiple correct data for the same input data is divided into multiple sets so that one input data corresponds to one correct data, and a classifier is created for each divided set Points to do.
[2] A different learning algorithm is applied to each divided set.
[3] A different learning parameter is applied to each divided set.
[4] Providing correct data of a plurality of evaluators, creating a set for each evaluator, and creating a plurality of classifiers corresponding to the evaluators.
[5] A point in which correct data of a plurality of evaluators is prepared, a set in which correct data of a plurality of evaluators is mixed, and a plurality of classifiers are generated.
An embodiment in which each of the above five features is provided, or an embodiment in which two or more features are arbitrarily selected and appropriately combined may be used.

  DESCRIPTION OF EMBODIMENTS Hereinafter, modes for carrying out the present invention will be described with some examples as follows.

Embodiment 1.
FIG. 1 shows a configuration example of a data classifier creating apparatus according to Embodiment 1 of the present invention.

In FIG. 1, the data classifier creating apparatus 1 according to the first embodiment includes a data storage unit 10, a one-to-one data creating unit 20, a classifier creating unit 30, and a classifier output unit 40. Yes. The data storage unit 10 further includes a case data storage unit 11, a one-to-one data storage unit 12, and a classifier storage unit 13. In FIG. 1, the data classifier creating apparatus 1 is configured as a software using a normal computer having a CPU, a memory, and a hard disk drive. You may make it comprise.

  FIG. 2 is a flowchart showing an example of the operation of the data classifier creating apparatus 1 shown in FIG.

  The operation of the data classifier creating apparatus 1 shown in FIG. 1 will be described with reference to the flowchart of FIG.

  First, the one-to-one data creation unit 20 creates a one-to-one data set that is a set of one-to-one data for creating a classifier based on the data stored in the case data storage unit 11. Stored in the one-to-one data storage unit 12 (step S10).

  Next, the classifier creating unit 30 creates a classifier based on the one-to-one data set stored in the one-to-one data storage unit 12, and stores it in the classifier storage unit 13 (step S20). Next, the classifier output unit 40 reads a plurality of classifiers from the classifier storage unit 13 and outputs the classifiers via the external connection unit 100 (step S30).

The data storage unit 10 is a large-capacity recording medium that can be accessed at high speed, such as a hard disk or a memory. The data storage unit 10 further includes a case data storage unit 11, a one-to-one data storage unit 12, and a classifier storage unit 13.

  The case data storage unit 11 stores a set of case data in which one or more correct data indicating a correct answer classification (class) is associated with one input data, that is, a case data set. The case data storage unit 11 also stores an evaluation ID for identifying an evaluator who gave correct data in association with the correct data. Input data is a set of data composed of continuous values or discrete values. For example, if the source of the data is acoustic data related to a song, it may be a feature value, tempo, genre number indicating a music genre, etc., with numerical values such as beat strength and frequency characteristics. For example, each pixel value of an image obtained by converting the image data to gray scale and reducing the size to 20 × 20 pixels. When the same evaluator gives a plurality of correct answer data, different evaluation IDs are prepared for each.

  FIG. 3 is a diagram illustrating an example of m pieces of case data stored in the case data storage unit 11.

  In FIG. 3, each case data stored in the case data storage unit 11 shows a state in which input data, correct answer data, and an evaluation ID given the correct answer data are stored in association with each other. For example, (C1, C1, C2) is assigned as correct data for input data (A11,..., A1n) having n variables, and an evaluation ID for identifying an evaluator who gave the correct data is , (M1, M2, F1) in order.

  Returning to FIG. 1, the one-to-one data creation unit 20 is a set of one-to-one data in which one input data corresponds to one correct data based on the case data set stored in the case data storage unit 11, that is, A plurality of one-to-one data sets are created, and the created one-to-one data sets are stored in the one-to-one data storage unit 12.

  The reason for creating a set of one-to-one data in which input data and correct answer data are one-to-one is that if a classifier is created with a plurality of correct answers in the same input data such as case data, This is because the classification accuracy decreases.

  FIG. 4 is a flowchart showing a processing example for the one-to-one data creation unit 20 to create a one-to-one data set for each evaluation ID.

  First, the one-to-one data creation unit 20 reads one case data from the case data storage unit 11 (step S110).

  Next, the one-to-one data creation unit 20 creates one-to-one data corresponding to the input data and the correct answer data one-to-one from the read case data by the number of correct answer data (step S120). . That is, when one-to-one data is created from the case data shown in FIG. 3 where the input data is {A11,..., A1n}, the input data {A11,..., A1n} and correct data {C1] and input data {A11 ,..., A1n} and correct answer data {C1}, input data {A11,..., A1n} and correct answer data {C2} are created.

  Next, the one-to-one data creation unit 20 stores each created one-to-one data in a one-to-one data set having the same identification name as the evaluation ID that gave the correct data of the one-to-one data (Step 1). S130).

  At this time, when the one-to-one data set corresponding to the evaluation ID does not exist, the one-to-one data creation unit 20 converts the one-to-one data set having the evaluation ID as an identification name into the one-to-one data storage unit 12. Create one-to-one data in the one-to-one data set.

  Next, the one-to-one data creation unit 20 determines whether or not all data has been read from the case data storage unit 11 (step S140). Here, if there is unread case data (“NO” in step S140), the process returns to step S110, and the processes after step S110 are performed. On the other hand, when the case data of all steps is read (step S140 “YES”), the one-to-one data creation unit 20 ends the above processing.

  Each of FIGS. 5A to 5C is a one-to-one storing a one-to-one data set composed of a one-to-one data label that is an identification name of a one-to-one data set and a plurality of one-to-one data. It is a figure which shows an example of the state of the data storage part.

  The one-to-one data label of the one-to-one data set is assigned the same identification name as the evaluation ID, and stores one-to-one data to which correct answer data is given by the same evaluation ID.

  FIG. 5A shows a one-to-one data set created with only the evaluation ID “M1”, and “M1” is assigned as a one-to-one data label.

  Similarly, FIG. 5B is a one-to-one data set created only with the evaluation ID “M2”, and FIG. 5C is a one-to-one data set created only with the evaluation ID “F1”. It is.

  The above is a method of creating a one-to-one data set for each evaluation ID by the one-to-one data creation unit 20. It is also possible to create a one-to-one data set in which evaluation IDs are mixed. In the case of creating a one-to-one data set in which evaluation IDs are mixed, the one-to-one data creation unit 20 performs a 1-to-1 data set between the same input data after the process of step S140 shown in FIG. A shuffle process for exchanging one-to-one data is performed.

FIG. 6 is a flowchart showing an example of the operation of the shuffle process of the one-to-one data creation unit 20.
First, the one-to-one data creation unit 20 selects a pair of one-to-one data sets that have not yet been shuffled from a plurality of one-to-one data sets created for each evaluation ID (step S151). .

  Next, the one-to-one data creation unit 20 selects one-to-one data having the same input data from a pair of the one-to-one data sets with a predetermined probability p (0.0 <p <1. 0) (step S152). This process is performed for all input data.

  Next, the one-to-one data creation unit 20 determines whether or not the shuffle process has been performed on all pairs of one-to-one data sets (step S420). If all have been completed (step S420 “Yes”), the shuffle process is terminated, and if there is a pair that has not been shuffled yet (step S420 “No”), the process returns to step S400. The above is the description of the shuffle process in step S150. The shuffle process may be repeated a plurality of times.

The above is a method for creating a one-to-one data set in which evaluation IDs are mixed. Other methods, for example, a predetermined number of empty one-to-one data sets are created in advance and created from case data. One-to-one data may be randomly stored in the one-to-one data set.
The above is the description of the one-to-one data creation unit 20.

Next, the classifier creation unit 30 will be described.
The classifier creation unit 30 creates a new classifier by applying a learning algorithm using the one-to-one data set created by the one-to-one data creation unit 20. A classifier classifies data into at least one class based on the data. Known techniques of learning algorithms include “decision tree”, “k-nearest neighbor method (k-NN)”, “neural network, support vector machine”, “Bayes model”, and the like. The learning algorithm is an algorithm for constructing a classifier itself based on learning data or an algorithm for selecting variables for learning data.

  FIG. 7 is a flowchart illustrating an example of an operation in which the classifier creating unit 30 creates a plurality of classifiers using a one-to-one data set.

  First, the classifier creation unit 30 reads the one-to-one data set group created by the one-to-one data creation unit 20 from the one-to-one data storage unit 12, and among the one-to-one data set group, One one-to-one data set that is not used as a learning data set is selected (step S21). The one-to-one data set group is a set of one-to-one data sets stored in the one-to-one data storage unit 12.

  In the example of the one-to-one data set in FIG. 5, the one-to-one data set “M1”, the one-to-one data set “M2”, and the one-to-one data set “F1” are collected. is there.

  Next, the classifier creating unit 30 selects a learning data set for use in construction of the classifier itself and an evaluation data set for use in evaluating the created classifier from the selected one-to-one data set. Select (step S22).

  Specifically, the classifier creation unit 30 equally divides the selected one-to-one data set into two, and selects one as a learning data set and the other as an evaluation data set. For example, the odd-numbered rows are used as learning data and the even-numbered rows are used as evaluation data, so that the data is equally divided into two. The evaluation data set may be selected from another one-to-one data set that is not a learning data set. In other words, a one-to-one data set that has not yet been selected as a learning data set is selected from the one-to-one data set group, and the evaluation data set is a one-to-one data set. A group other than the one-to-one data set selected as the learning data set is randomly selected from the group.

  An example of selecting the learning data set and the evaluation data set from the one-to-one data set group shown in FIG. 5 shows that when the one-to-one data set “M1” is selected as the learning data set, the remaining two One of the one-to-one data sets “M2” and “F1” is selected at random, and the selected one-to-one data set is set as an evaluation data set.

  Next, the classifier creation unit 30 selects a classifier to be newly created, a construction parameter for creating the classifier, and a learning algorithm (step S23). Therefore, the classifier creation unit 30 holds a classifier pair that combines a classifier, a construction parameter, and a learning algorithm in advance. The construction parameters are functions used for construction of the classifier, threshold values, and the like.

  FIG. 8 shows an example of a classifier pair. A classifier, a construction parameter of the classifier, and an algorithm for learning are combined, and a classifier pair ID that is an identifier is allocated and stored in each pair. An example of a state is shown.

  The classifier creation unit 30 selects one classifier pair in which the classifier is not created from the currently selected learning data set in the order of the classifier pair ID. If no construction parameters are specified, the classifier is built with the most standard construction parameters.

  Further, in FIG. 8, the classifier pair IDs “AlgoID1” and “AlgoID2” are both the same classifier k-NN, and the learning algorithm is equally a genetic algorithm, but the construction parameter is “k”, respectively. = 3 "and" k = 5 ". Even the same classifier may be treated as another classifier if the construction parameters are different.

  In FIG. 8, the classifier pair IDs “AlgoID2” and “AlgoID3” are both similar classifiers k-NN, but the learning algorithms are “genetic algorithm” and “principal component analysis”, respectively. Is different. Even if they are the same classifier, they may be treated as different classifiers if the learning algorithms are different.

  Next, the classifier creating unit 30 creates a new classifier based on the learning data set selected in step S22, the evaluation data set, and the classifier pair selected in step S23 (step S23). S24). If classifiers have been created for all classifier pairs in the selected learning data set (step S25 “YES”), the process proceeds to step S26; otherwise (step S25 “NO”). The process returns to step S23 (step S25).

  Next, the classifier creation unit 30 stores the created classifier in the classifier storage unit 13 (step S26). The classifier storage unit 13 combines the identifier of the one-to-one data set used as the learning data set, the correct answer rate that is the ratio that the classification result of the evaluation data matches the correct answer data, and the classifier construction information. Store. Here, the classifier construction information is information necessary to reconstruct the created classifier. For example, classifier types such as decision trees and k-NN, threshold values, weights, teacher data, branch conditions, and the like are stored.

  FIG. 9 is a diagram illustrating an example of six classifiers stored in the classifier storage unit 13.

  In FIG. 9, as information for one classifier, the identifier of the created classifier (classifier identifier), the identifier of the one-to-one data set used as the learning data set, the correct answer rate, and the classifier construction information It shows a state of storing in combination. For example, the classifier identifier “classifier 1” indicates that the identifier of the one-to-one data set is “M1” and the correct answer rate is “0.80”.

  Further, the classifier name “k-NN” and parameters necessary for constructing the classifier are stored as classifier construction information.

  Next, the classifier creating unit 30 determines whether or not all the one-to-one data sets have been used as learning data sets (step S27). Here, when there is a one-to-one data set that has not been used as a learning data set (step S27 “NO”), the classifier creation unit 30 returns to step S21, but if all have been completed (step S27). (S27 “YES”), the classifier creation unit 30 ends the process.

  Here, as a specific example of the classifier creation process in step S24, a method of creating “k-NN” which is one of the classifiers and performing variable selection using a genetic algorithm (GA) as a learning algorithm is shown. .

  “K-NN” means that when assigning a class to a node X whose class is unknown, the class of k nodes is referred to in order of the distance from the node X, and the most frequently appearing class is defined as the class of the node X. It is a method to do. The distance between the nodes is obtained by calculating the Euclidean distance between the input data, but instead of the Euclidean distance, a Mahalanobis general distance or the like may be used.

  By the way, if all the variables of the input data are used for calculating the distance, the classification accuracy may be lowered. For example, if the variable includes noise data that hardly affects the classification, or if a duplicate event is described, the accuracy of the classification decreases. Therefore, it is necessary to select which variable is used. The combination is 2 to the Nth power, assuming that there are N variables. In other words, if all combinations are evaluated on a computer that takes 1 millisecond to evaluate one combination, the processing will be completed in about 17 minutes if there are at most 20 variables in the input data. If the process is not continued for tens of thousands of years, it will not be terminated.

  Therefore, when the number of variables is large to some extent, it is necessary to apply a learning algorithm to select which variable to use. Therefore, a genetic algorithm (GA) is used as the variable selection learning algorithm. The detailed algorithm of GA is disclosed in "Genetic Algorithms in Searching, Optimization, and Machine Learning" by David E. Goldberg.

  Genetic algorithm (GA) is an algorithm inspired by biological evolution, such as genetic selection, crossover, mutation, and generational change. In the genetic algorithm (GA), a problem to be solved is expressed as a gene by a bit string.

  FIG. 10 shows an example in which the variable selection problem is expressed as a gene (bit string).

  FIG. 10 shows a state in which [1: used, 0: not used] is set for n variables (A1,..., An) for Z gene individuals. The classifier creating unit 30 initially sets the initial population of these Z gene individuals at random. The number of genes Z is the number of gene individuals per generation. There is no particular regulation regarding the Z value, but 5 or more is desirable.

  FIG. 11 is a flowchart illustrating an example of a learning procedure using a genetic algorithm (GA). As shown in FIG. 11, the process from step S32 for executing the evaluation function to step S37 for performing the generation change process is repeatedly executed. This iterative process is called a generation in the genetic algorithm (GA).

  First, the classifier creation unit 30 generates Z initial gene individuals (step S31). Each bit of the gene individual is randomly generated.

  Next, the classifier creating unit 30 performs “k-NN” on the evaluation data set based on the setting of each of the Z gene individuals, and calculates the fitness (step S32). The fitness is calculated by actually classifying the evaluation data using “k-NN” created based on the learning data set. The learning data set is T, the evaluation data set is E, the i-th gene value is Gi, the i-th variable in the x-th learning data is t (x, i), and the y-th evaluation data is Assuming that the i-th variable is E (y, i), the Euclidean distance D (x, y) between the learning data x and the evaluation data y is obtained by the following equation (1).

  The classifier creating unit 30 performs this distance calculation for all learning data with x∈T. After calculating the distance to all the learning data, k learning data are acquired in ascending order of the calculated Euclidean distance D (x, y), and the mode value of the correct data is calculated. It is checked whether or not the mode value matches the correct answer data corresponding to the y-th evaluation data. If they match, the correct answer number c is incremented by one. The above processing is performed on all the evaluation data for which yεE, and the correct answer rate with the correct answer data is calculated. That is, the correct answer rate Rj of the j-th gene individual is obtained by the following equation (2), where M is the total number of evaluation data.

  The classifier creation unit 30 sets the correct answer rate Rj obtained here as the fitness of the gene individual j. This fitness calculation is performed for all Z gene individuals.

  Next, the classifier creating unit 30 determines whether or not the maximum adaptability obtained in step S32 is the maximum adaptability in the previous generations (step S33). If the fitness level is the maximum or the first generation so far (step S33 “YES”), the process proceeds to step S34. Otherwise (step S33 “No”), the process proceeds to step S35.

  Next, the classifier creation unit 30 associates the bit sequence of the gene individual that has obtained the best fitness with the fitness, and temporarily stores it in the storage unit 20 (step S34).

  Next, the classifier creation unit 30 determines whether to end learning (step S35). Learning ends when the number of generations exceeds a predetermined threshold, when the maximum fitness exceeds a predetermined threshold, or the average fitness of all gene individuals exceeds a specified threshold. Judgment in case of All conditions may be judged, or only one of them may be judged. If it is determined to end the learning (step S35 “YES”), the classifier creation process is ended. Otherwise (step S35 “No”), the process proceeds to step S36.

  Next, the classifier creation unit 30 performs a process of setting a next-generation gene individual (step S36). This process is an essential part of the genetic algorithm (GA), and specifically, crossover and mutation of gene individuals and selection of gene individuals to be left in the next generation are performed.

  FIG. 12 is a flowchart showing an example of the operation of a general generation change process in the genetic algorithm (GA).

  First, the classifier creation unit 30 selects a parent pair in order to create a child gene individual from a parent gene pair (step S40). Next, the classifier creation unit 30 crosses the parent pair of gene individuals selected in step S40 to create a child gene individual (step S41). Next, the classifier creation unit 30 performs a mutation process on the created child gene individual (step S42). Next, the classifier creation unit 30 performs survival selection as to which gene individual is left in the next generation among the generations of the parent (step S43). The above is the flow of the generation change process. There are various methods for replication selection, crossover, mutation, and survival selection, and any method or combination may be used.

  Here, the generation change process will be described by taking a method called a simple GA model as an example. In the simple GA model, roulette selection is performed as replication selection. Roulette selection is a method of selecting a gene individual with a probability proportional to the fitness of the gene individual. The selection probability p (j) of the gene individual j is obtained by the following equation (3).

  The classifier creation unit 30 selects a gene individual probabilistically using random numbers so as to follow p (j) obtained here. Since a selected gene individual is paired to create a child gene individual, at least two are selected.

  Next, the classifier creation unit 30 performs one-point crossover to create a child gene individual using the gene pair selected in step S40. One-point crossing is a method of designating one arbitrary cutting position of the bit sequence and crossing the parent gene pair at that position.

  FIG. 13 shows a case where a gene individual A having a bit string of {100101} and a gene individual B having a bit string of {000111} are set as a parent pair, and a cut point is designated at the third bit and one point crossover is performed. This shows how gene individuals C and D are created.

  That is, the bit string of the gene individual A to the left of the cut position is A1 = {100}, the right bit string is A2 = {101}, the bit string of the gene individual B to the left of the cut position is B1 = {000}, and the right bit string is If B2 = [111], the bit string of gene individual C newly created by one-point crossover is {A1} {B2} = {100111}, and the bit string of gene individual D is {B1} {A2} = {000101} It shows how it becomes.

  Next, an example of mutation processing in a simple GA model is shown. For each bit of a newly created gene individual, the bit is inverted at a predetermined rate. The occurrence probability of bit inversion is set to a small value of about 0.1%.

  Next, as for survival selection in a simple GA model, only a newly created gene individual is left in the next generation, and all the old generation gene individuals are deceived. For the number of genes less than one generation, a bit string of gene individuals is randomly generated and newly created.

  The above is an example of the generation change process. When the generation change process is completed, the process returns to step S32, and the fitness is calculated for the new generation of gene individuals.

  The above is an example of creating “k-NN” by applying GA to the learning algorithm.

  Although the creation process of “k-NN” has been described in detail, it is needless to say that a classifier can be created using another classifier or a learning algorithm. For example, a decision tree that is one of the classifiers can be created. In the decision tree, CART, C4.5, and the like are known as learning algorithms. A decision tree can be created according to these learning algorithms. In addition, a decision tree can be constructed by applying GA to selection of variables for learning data.

  A neural network such as a multilayer perceptron can also be created using the learning data set and the evaluation data set.

  A support vector machine (SVM) can also be used as a classifier. In SVM, in order to cope with a non-linear separation problem, non-linear SVM using a kernel function is known. As a kernel function used at this time, a polynomial type kernel, a Gaussian type kernel, or the like is known, but an SVM may be created using these as different construction parameters.

  Further, the classifier may be created by performing normalization processing of the input data.

  In this embodiment, the variable used in “k-NN” is selected using a genetic algorithm, but the present invention is not limited to this, and other types of classifications such as decision trees and support vector machines are used. When creating a container, the variables constituting the input data may be selected using a genetic algorithm.

  The classifier output unit 40 selects a predetermined number of classifiers stored in the classifier storage unit 13 in descending order of the correct answer rate, and outputs the selected classifiers via the external connection unit 100. For example, in the example of the classifier stored in the classifier storage unit 13 illustrated in FIG. 9, when selecting the top three classifiers with the highest correct answer rate, “Classifier 5” with a correct answer rate of 0.85 and the correct answer “Classifier 1” with a rate of 0.80 and “Classifier 6” with a correct answer rate of 0.75 are selected.

  Alternatively, a predetermined number of classifiers created from the same learning data set may be acquired in descending order of the correct answer rate. For example, in the example of FIG. 9, there are three types of identifiers of the one-to-one data set: “M1”, “M2”, and “F1”. For each of these identifiers, when selecting a classifier with a high correct answer rate, a classifier created from “M1” is created from “M2”, “Classifier 1” with a correct answer rate of 0.80. The classifier selects “Classifier 4” with a correct answer rate of 0.68, and the classifier created from “F1” selects “Classifier 5” with a correct answer rate of 0.85.

  In addition, a classifier having a higher correct answer rate than a predetermined value may be selected. For example, when a classifier having a correct answer rate of 0.80 or more is selected, in the example of FIG. 9, a classifier 1 having a correct answer rate of 0.80 and a classifier 5 having a correct answer rate of 0.85 are selected. .

  The number of classifiers to be output is based on a predetermined number that is predetermined in the classifier output unit 40, but the number specified from the outside may be output. Alternatively, all classifiers may be output.

  The classifier selected by the classifier output unit 40 is output via the external connection unit 100. The external connection unit 100 is a data communication path, and is a bus, a network cable, or the like.

  The above is the description of the data classifier creating apparatus 1.

  Therefore, according to the data classifier creating apparatus 1 of the first embodiment, since a plurality of classifiers are created and stored, a plurality of evaluators or a plurality of judgment criteria can be handled for one input data. Digital content can be classified by handling a plurality of evaluators or a plurality of judgment criteria for one input data.

  As a result, even if a classifier is constructed with a data set in which a plurality of correct answer data are assigned to one input data, a classifier with high classification accuracy can be created.

  In the data classifier creating apparatus 1 according to the first embodiment, even when a plurality of correct answer data are given to a single input data by a plurality of evaluators, the input data and the correct answer data are one. Since the classifier is created by the one-to-one data set that collects the one-to-one data that is one-to-one, it is possible to create a classifier with high accuracy, which is not biased toward the evaluation of a specific evaluator. A good classifier can be created.

Embodiment 2. FIG.
FIG. 14 is a block diagram showing a configuration example of the data classification device 2 according to Embodiment 2 of the present invention.

  14, the data classification device 2 according to the second embodiment includes a classifier acquisition unit 60, a classifier storage unit 70, a classification execution unit 80, and a classification result integration unit 90. The data classification device 2 is connected to the data classifier creation device 1 according to the first embodiment shown in FIG. The data classification device 2 may also be configured as software, similar to the data classification device 2 of the first embodiment shown in FIG.

  FIG. 15 is a flowchart showing an example of the operation of the data classification device 2 shown in FIG.

  Referring to FIG. 15, the operation of data classification device 2 in the second embodiment shown in FIG. 14 will be described.

  First, the classifier acquisition unit 60 of the data classification device 2 acquires a classifier from the data classifier creation device 1 via the external connection unit 100, and stores the acquired classifier in the classifier storage unit 70 (step). S200).

  Next, the classification execution unit 80 of the data classification device 2 determines whether or not input data has been input from the external device 200 (step S210). If input data has been input (step S210 “YES”), the process proceeds to step S220. If not (step S210 “NO”), the process waits until input data is input. The external device 200 is connected to a cable such as a bus or LAN, or an input device such as a mouse.

  Next, the classification execution unit 80 of the data classification device 2 classifies the input data that has been input using the classifier stored in the classifier storage unit 70. (Step S220).

  Next, the classification result integration unit 90 of the data classification device 2 integrates the classification results output from the plurality of classifiers, and determines the final classification result (step S230).

  Next, the classification execution unit 80 of the data classification device 2 outputs the final classification result to the external device 200 that has input the data (step S240). The above is the outline of the operation of the data classification device 2.

  Next, each part of the data classification device 2 will be described in detail.

  The classifier acquisition unit 60 acquires the number of classifiers set in the data classifier generation device 1 from the data classifier generation device 1 via the external connection unit 100, and stores the acquired classifiers as classifiers. Store in the unit 70. The classifier storage unit 70 stores the same data as the classifier storage unit 13 of the data classifier creation device 1. That is, as shown in FIG. 9, the identifier of the one-to-one data set used as the learning data set, the correct answer rate, and the classifier construction information are stored in combination.

  The classifier acquisition unit 60 may specify the number of classifiers to be acquired in the data classifier creation apparatus 1 and acquire the same number of classifiers. Further, all classifiers may be acquired from the data classifier creating apparatus 1. Further, a classifier having a correct answer rate equal to or higher than a predetermined value may be acquired from the data classifier creating apparatus 1.

  When input data is input from the outside, the classification execution unit 80 gives input data to a plurality of classifiers stored in the classifier storage unit 70, and obtains a plurality of classification results. For example, when obtaining classification results by providing input data to six classifiers, a total of six classifications, one from each classifier, such as {C1, C1, C1, C2, C2, C3}. Get the result.

  In the classification example {C1, C1, C1, C2, C2, C3}, there are three unique classes C1, C2, and C3. Therefore, in order to make the final classification result one class, It is necessary to integrate the classification results into one class.

  The classification result integration unit 90 integrates a plurality of classification results obtained by the classification execution unit 80, and determines one class as a final classification result. The classification result integration unit 90 uses a majority voting method that selects, from the classification results acquired from the classifiers, the class output from the largest number of classifiers as the final classification result. For example, it is assumed that the classification result is {C1, C1, C1, C2, C2, C3}. When the majority process is performed in the example of the classification result, C1 has three votes, C2 has two votes, and C3 has one vote. Therefore, C1 that has acquired the maximum number of votes is selected as the final classification result.

  Also, the number of votes may be calculated by weighting one vote of the classifier, such as setting the size of one vote of each classifier as the value of the correct answer rate. For example, when a classifier with a correct answer rate of 0.8 and a classifier with a correct answer rate of 0.7 output the same class as a classification result, 1.5 is obtained by adding 0.8 and 0.7. , The number of votes for that class.

  The above is the description of the data classification device 2.

  Therefore, according to the data classification device 2 of the second embodiment, a plurality of classifiers created and stored by the data classifier creation device 1 of the first embodiment are acquired, and classification is performed by the plurality of classifiers. By integrating the classification results, it is possible to select the optimum final classification result, and as a result, it is possible to obtain a classification result with higher accuracy than in the past.

Embodiment 3 FIG.
The data classifier creating apparatus 3 according to the third embodiment is not limited to the same classifier (general classifier) as that created by the data classifier creating apparatus 1 according to the first embodiment, but only limited types of correct answer data. A plurality of classifiers (dedicated classifiers) are also created from the partial one-to-one data set configured as described above. Since the dedicated classifier is created from a partial one-to-one data set composed only of limited types of correct answer data, the output data is also limited to the same type of class as the correct answer data used for creation. The Such a dedicated classifier has a higher classification to the final class when the result of classification by the general classifier does not show a difference of more than a predetermined vote between classes with the highest number of votes. Used for final voting for accuracy. For example, when the final voting is performed for two classes C1 and C2, a dedicated classifier created based on a one-to-one data set including only C1 and C2 as correct answer data is used. Since this dedicated classifier is a classifier created from a one-to-one data set including only C1 and C2 as correct data, the classification result to be output is always one of classes C1 and C2.

  FIG. 16 is a block diagram illustrating a configuration example of the data classifier creating apparatus 3 according to the third embodiment.

  In FIG. 16, the data classification creation device 3 has the same configuration as the data classifier creation device 1, but the content stored in the one-to-one data storage unit 12, the classifier storage unit 13, and the one-to-one data creation. Since the operation contents of the unit 20 and the classifier creation unit 30 are different, in the third embodiment, the one-to-one data storage unit 12A, the classifier storage unit 13A, the one-to-one data creation unit 20A, and the classifier creation, respectively. Part 30A. The data classifier creating apparatus 3 may also be configured as software, similar to the data classifier creating apparatus 1 of the first embodiment.

  Here, the one-to-one data creation unit 20A of the data classifier creation device 3 according to the third embodiment sets specific correct answer data in addition to the one-to-one data set creation processing of the data classifier creation device 1. One or more partial one-to-one data sets composed of only one-to-one data are created and stored in the one-to-one data storage unit 12A. This partial one-to-one data set is based on the fact that a combination of two types of correct answer data is used as a child element and the number of combinations is created. Note that the number is not limited to two, but may be a number larger than two. Alternatively, the designer may specify in advance a combination for creating a partial one-to-one data set without creating a partial one-to-one data set for all combinations.

  FIG. 17 is a flowchart showing an example of the operation of the one-to-one data creation unit 20A. Similar steps are assigned to steps that perform the same processing as the one-to-one data creation unit 20.

  First, the one-to-one data creation unit 20A creates an empty partial one-to-one data set having two types of classes as components in the one-to-one data storage unit 12 (step S100). The partial one-to-one data set is created by the unique number of evaluation IDs for all combinations of two types of classes. For example, when there are 10 combinations of classes and the unique number of evaluation IDs is 3, 30 partial one-to-one data sets are created with 10 × 3 = 30. In addition, a combination of a component and an evaluation ID is given as an identifier of each partial one-to-one data set. For example, when the component classes are “C1” and “C2” and the evaluation ID is “M1”, “M1 {C1, C2}” or the like is given as the identifier of the partial one-to-one data set.

  Subsequent step S110, step S120, and step S130 are the same as the one-to-one data creation unit 20 of the first embodiment.

  Next, the one-to-one data creation unit 20A adds one-to-one data to the corresponding partial one-to-one data set based on the created correct data of the one-to-one data and the evaluation ID (Step 1). S135). When there are a plurality of corresponding partial one-to-one data sets, one-to-one data is added to all the corresponding partial one-to-one data sets.

  FIGS. 18A to 18C show a state in which one-to-one data is stored in a partial one-to-one data set having only classes C1 and C2 for each evaluation ID “M1, M2, F1”. It is shown.

  The three partial one-to-one data sets shown in FIG. 18 store only the one-to-one data to which the correct data C1 and C2 are allocated. FIG. 18A shows an evaluation ID “M1”. FIG. 18B stores information assigned by the evaluation ID “M2”, and FIG. 18C stores information assigned by the evaluation ID “F1”.

  The next step S140 is the same as the one-to-one data creation unit 20 of the data classifier creation device 1. In addition to the one-to-one data set shown in FIG. 5, the partial one-to-one data set shown in FIG. 18 is also stored in the one-to-one data storage unit 12A by the above-described one-to-one data creation unit 20A.

  Similarly to the one-to-one data creation unit 20 of the data classifier creation device 1, a partial one-to-one data set in which evaluation IDs are mixed can be created by performing a shuffle process. The shuffle process is performed after step S140, as in the one-to-one data creation unit 20. The flow of operation is almost the same as the flowchart shown in FIG. 6, but the processing in step S151 is slightly different. Specifically, in order to mix evaluation IDs of partial one-to-one data sets, a pair of partial one-to-one data sets for exchanging one-to-one data is selected from those that have not yet been shuffled. Assume that pairs of partial one-to-one data sets have the same combination of correct data in one-to-one data and different evaluation IDs. Other steps S152 and S153 are the same as those in the first embodiment.

  The above is the description of the one-to-one data creation unit 20A.

  The operation flow of the classifier creating unit 30A is the same as that of the classifier creating unit 30 of the data classifier creating apparatus 1 shown in FIG. 7, but the contents of the operations in step S21, step S22, and step S26 are different.

  First, the classifier creation unit 30A selects one one-to-one data set that has not yet been selected from the one-to-one data set group stored in the one-to-one data storage unit 12A. One partial one-to-one data set is also selected from one data set group (step S21). The partial one-to-one data set group is a collection of partial one-to-one data sets having the same components. For example, the partial one-to-one data set M1 {C1, C2}, the partial one-to-one data set M2 {C1, C2}, and the partial one-to-one data set M3 {C1, C2} shown in FIG. Since the partial one-to-one data set having the constituent elements {C1, C2} is a partial one-to-one data set group. There are as many partial one-to-one data sets as there are combinations of components, and one partial one-to-one data set is selected for each partial one-to-one data set group.

  Next, the classifier creating unit 30A selects a learning data set and an evaluation data set from the one-to-one data set selected in step S21 and the partial one-to-one data set (step S22). The processing for selecting the learning data set and the evaluation data set from the one-to-one data set is the same as in the first embodiment.

  The classifier creating unit 30A further selects a learning data set and an evaluation data set from the partial one-to-one data set. In this case, the partial one-to-one data set selected in step S21 is equally divided into two, and one is set as a learning data set and the other is set as an evaluation data set. For example, if odd lines are learning data and even lines are evaluation data, the data can be equally divided into two. This process is performed on all selected partial one-to-one data sets.

  Further, the evaluation data set of the partial one-to-one data set was selected as the learning data set from the partial one-to-one data sets belonging to the same partial one-to-one data set group as the learning data set. Other partial 1-to-1 data sets may be selected at random.

  Subsequent step S23, step S24, and step S25 are the same as in the first embodiment.

  Next, the classifier creation unit 30A stores the created classifier in the classifier storage unit 13A (step S26). The classifier creating unit 30A includes a general classifier that is a classifier that creates a one-to-one data set as learning data, and a dedicated classifier that is a classifier that creates a partial one-to-one data set as learning data. Separately, it is stored in the classifier storage unit 13A.

  FIG. 19 is an example of a classifier stored in the classifier storage unit 13A. FIG. 19A is a storage example of a general classifier, and FIG. 19B is a storage example of a dedicated classifier. The format of the general classifier in FIG. 19A is the same as that in FIG. 9 described in the first embodiment. FIG. 19B stores classifier identifiers, partial one-to-one data set identifiers, components, correct answer rates, and classifier construction information. The classifier identifier, the correct answer rate, and the classifier construction information are the same as those of the general classifier of (a). The identifier of the partial one-to-one data set is the identifier of the learning data set used for creating the dedicated classifier. The component indicates a component of the partial one-to-one data set used for the construction of the dedicated classifier. For example, in the three dedicated classifiers shown in FIG. 19B, the components {C1, C2} are stored. This indicates that it is a dedicated classifier for classifying C1 and C2 created from a partial one-to-one data set of only one-to-one data whose correct answer data is C1 and C2.

The next step S27 is the same as that in the first embodiment.
The above is the description of the data classifier creation device 3.

  Therefore, according to the data classifier creating apparatus 3 of the third embodiment, a plurality of classifiers (general classifiers) similar to those created by the data classifier creating apparatus 1 of the first embodiment can be created. At the same time, a plurality of classifiers (dedicated classifiers) can be created from a partial one-to-one data set composed of only some classes of correct data.

Embodiment 4 FIG.
FIG. 20 is a block diagram showing a configuration example of the data classification device 4 according to Embodiment 4 of the present invention.

  20, the data classification device 4 includes a classifier acquisition unit 60A, a classifier storage unit 70, a classification execution unit 80A, and a classification result integration unit 90, and is illustrated in FIG. 16 via the external connection unit 100. It is connected to the data classification creation device 3 of the third embodiment. Since the classifier storage unit 70 and the classification result integration unit 90 are the same as those of the data classification device 2 in the second embodiment, the description thereof is omitted.

  Here, in the data classification device 4 according to the fourth embodiment, the operations of the classifier acquisition unit 60 and the classification execution unit 80 of the data classification device 2 according to the second embodiment are different. 60A and classification execution unit 80A. The classifier acquisition unit 60 of the data classification device 4 is connected to the data classifier creation device 3 through the external connection unit 100. The classification execution unit 80 is connected to the external device 200. This data classification device 4 may also be configured as software, similar to the data classification devices 1 to 3 of the first to third embodiments.

  Similar to the classifier acquisition unit 60 of the data classifier 2, the classifier acquisition unit 60A in the data classifier 4 acquires a general classifier from the data classifier creation device 3, and the dedicated classifier also includes a data classifier. Processing to be acquired from the creation device 3 is performed.

  Similar to the classification execution unit 80 of the data classification device 2, the classification execution unit 80A in the data classification device 4 performs classification using a plurality of general classifiers, and the classification result of the general classifiers satisfies a predetermined condition. If not, the classifier acquisition unit 60A acquires a dedicated classifier, and performs a voting vote using the dedicated classifier.

  FIG. 21 is a flowchart showing an example of the operation of the data classification device 4.

  First, the classifier acquisition unit 60A of the data classification device 4 acquires a general classifier from the data classifier creation device 3 via the external connection unit 100 and stores it in the classifier storage unit 70 (step S200). This process is the same as that of the classifier acquisition unit 60 of the second embodiment.

  The next step S210 is the same as in the second embodiment.

  Next, the classification execution unit 80A performs classification based on the general classifier acquired in Step S200 (Step S220). This process is the same as that of the classification execution unit 80 of the second embodiment.

  Next, the classification execution unit 80A determines whether or not the number of votes obtained for the class that has obtained the maximum number of votes exceeds a predetermined value (step S221). If it exceeds the predetermined value, the process proceeds to step S230, and if not, the process proceeds to step S222.

  Next, when the maximum number of votes does not exceed a predetermined value, the classifier acquisition unit 60A acquires the dedicated classifier created from the two classes with the highest number of votes from the data classifier creation device 3 (step S222). . For example, when the two classes with the highest number of votes are C1 and C2, a dedicated classifier created from a learning data set having only C1 and C2 as constituent elements is acquired. The dedicated classifiers also acquire the dedicated classifiers by the number set in the data classifier creation device 3 in the same manner as the general classifier acquisition process. Alternatively, the number of dedicated classifiers to be acquired may be designated and the number of dedicated classifiers acquired.

  Next, the classification execution unit 80A performs classification again on the same input data with the dedicated classifier acquired as described above (step S223). That is, in step S223, the classification execution unit 80A performs the same process as step S220, except that the classifier that performs classification is a dedicated classifier. If classification is performed by the dedicated classifiers C1 and C2, the classification result of either C1 or C2 is output by the dedicated classifier.

  Next, the classification result integration unit 90 integrates the classification results, determines the final classification result, and returns it to the classification execution unit 80A (step S230). This process is the same as in the second embodiment.

  The next step S230 is the same as in the second embodiment.

  The above is the operation flow of the data classification device 4 of the fourth embodiment.

  In step S200, the classifier acquisition unit 60A acquires all the dedicated classifiers and stores them in the classifier storage unit 70. From the dedicated classifier stored in the classifier storage unit 70 in step S222, the classifier acquisition unit 60A acquires all the dedicated classifiers. The classifier required for the final vote may be read out.

  Further, although the final vote in the two classes with the highest number of votes, the final vote in two or more classes, such as a final vote by a class exceeding a predetermined vote rate, may be performed.

  Therefore, according to the data classification device 4 of the fourth embodiment, the same effects as those of the data classification device 2 of the second embodiment can be obtained, and further, the data classification device 4 can be created by the data classifier creation device 3 of the third embodiment. By using a dedicated classifier that specially classifies specific classes, the classification accuracy can be further improved.

DESCRIPTION OF SYMBOLS 1,3 Data classifier preparation apparatus 2,4 Data classification apparatus 10 Storage part 11 Case data storage part 12, 12A One-to-one data storage part 13, 13A Classifier storage part 20, 20A One-to-one data preparation part 30, 30A Classifier creation unit 40 Classifier output unit 60, 60A Classifier acquisition unit 70 Classifier storage unit 80, 80A Classification execution unit 90 Classification result integration unit 100 External connection unit 200 External device

Claims (11)

Case data in which input data corresponds to correct data indicating a correct answer classification of the input data, and includes cases in which a plurality of correct data given by a plurality of evaluators corresponds to the same input data A case data storage for storing a set of case data;
One-to-one data creation that reads a set of case data from the case data storage unit and creates a plurality of one-to-one data sets corresponding to one correct data for each of the plurality of evaluators. And
Classifier creation for creating a plurality of classifiers by controlling to create one classifier from one set of one-to-one data while referring to the plurality of sets of one-to-one data And
A data classifier creating apparatus having   2. The data classifier according to claim 1, wherein the classifier creating unit creates the plurality of classifiers by applying two or more types of learning algorithms to the plurality of created one-to-one data sets. Creation device.   The classifier creating unit creates the plurality of classifiers by applying a learning algorithm using two or more different learning parameters to the created plurality of one-to-one data sets. The data classifier preparation apparatus of Claim 1 or Claim 2. The data classifier according to any one of claims 1 to 3 , wherein the classifier creation unit further creates a classifier by selecting a variable constituting input data using a genetic algorithm. Device making device. A classifier storage unit that stores a plurality of classifiers created by the data classifier creation device according to any one of claims 1 to 4,
A classification execution unit that inputs input data to each of the plurality of classifiers stored in the classifier storage unit to obtain a plurality of classification results;
A classification result integration unit that outputs a final class using the plurality of classification results ;
A data classification device. The classifier storage unit, the classification result integration unit of the class to be output target, all types of classes and can output dedicated classifier only some types of classes, the classification result integration unit is to output target storing an output capable general classifier class,
The data classification device according to claim 5 . The classifier storage unit further stores a correct answer rate of classification results corresponding to the plurality of classifiers,
The classification result integration unit weights the plurality of classification results according to the correct answer rate and outputs the final class;
The data classification device according to claim 5 or 6. Case data in which input data corresponds to correct data indicating a correct answer classification of the input data, and includes cases in which a plurality of correct data given by a plurality of evaluators corresponds to the same input data Reading a set of case data from a set of case data, creating a plurality of sets of one-to-one data in which one input data corresponds to one correct data for each of the plurality of evaluators ;
Controlling to create one classifier from one one-to-one data set while referring to the plurality of one-to-one data sets created, and creating a plurality of classifiers;
A method for creating a data classifier. Storing a plurality of classifiers created by the data classifier creation method according to claim 8;
Inputting input data to each of the plurality of classifiers stored in the classifier storage step to obtain a plurality of classification results;
Using the plurality of classification results to output a final class;
A data classification method comprising: Case data in which input data corresponds to correct data indicating a correct answer classification of the input data, and includes cases in which a plurality of correct data given by a plurality of evaluators corresponds to the same input data Reading a set of case data from a set of case data, creating a plurality of one-to-one data sets for each of the plurality of evaluators, one input data corresponding to one correct data;
Controlling to create one classifier from one one-to-one data set while referring to the plurality of one-to-one data sets created, and creating a plurality of classifiers;
Data classifier creation program that causes a computer to execute. Storing a plurality of classifiers created by the data classifier creation program according to claim 10;
Inputting input data to each of the plurality of classifiers stored in the classifier storage step to obtain a plurality of classification results;
Using the plurality of classification results to output a final class;
Classification program that causes a computer to execute.

Images