JP5733229B2 - Classifier creation device, classifier creation method, and computer program - Google Patents

Description

  The present invention relates to a classifier creating apparatus, a classifier creating method, and a computer program. In particular, a classifier for determining which of two classes a given data belongs to is learned with a known class. It is suitable for use to automatically construct using data.

  Conventionally, a classifier, which is a device for determining which of two classes a given data belongs to based on a certain rule, has been used in various industries. In the manufacturing industry, for example, a classifier is used to automatically determine whether a product is a non-defective product from measurement data of the product by a sensor. In the medical field, for example, a classifier is used to automatically determine from a test result whether or not there is a high suspicion of a certain disease.

  Automatic creation of rules in such a classifier using a large number of learning data whose classes are known is called supervised learning. When creating a classifier by supervised learning, a highly accurate classifier is often not created if the composition ratio of the class to which the learning data belongs is extremely biased. For example, if a rule for judging whether or not a product is non-defective from the measurement data of the product by the sensor is created so that the correct answer rate (the ratio that the predicted class matches the actual class) is high, it is generally not possible. Since the number of non-defective products is extremely smaller than the number of non-defective products, a classifier that always predicts a class to which a relatively large number of learning data belongs (non-defective product), or a class to which a relatively large number of learning data belongs (non-defective product). A classifier with an extremely high frequency of prediction may be generated. In the following description, “a class to which a relatively small number of learning data belongs” is referred to as a “minority class” as necessary, and “a class to which a relatively large number of learning data belongs” is referred to as a “major class” as necessary. Called.

  Thus, when creating a classifier using learning data in which the class composition ratio is extremely biased, it is determined that data that should belong to a minority class belongs to a majority class (so-called minority class There are many cases where it is desirable to avoid as much as possible. For example, a case where it is acceptable to judge a non-defective product as a non-defective product to some extent, but a case where it is desired to avoid judging a defective product as a non-defective product corresponds to such a case.

However, as described above, the classifier rule learning algorithm is usually made so that the accuracy rate is as high as possible. Since the failure of a small number of classes and the loss of a large number of classes are considered on the same scale, it is often impossible to meet the demand as in the case described above and a high-precision classifier cannot be obtained.
Thus, learning data is processed (sampled) to create a classifier. Specifically, at least a part of the learning data of the minority class is copied so that the number of learning data of the minority class is N times the number of learning data of the minority class (N> 1). There is a method called an oversampling method in which the number of learning data is increased and the classifier rules are created using the small number of learning data and the large number of learning data thus increased. Contrary to the oversampling method, a part of the learning data of many classes is arranged so that the number of learning data of many classes is 1 / M times the number of learning data of the many classes (M> 1). This is called an undersampling method that creates a classifier rule by using a large number of learning data and a small number of learning data that have been reduced in this way to reduce the number of learning data for many classes. There is a way to be. Here, the aforementioned N and M are referred to as “weights”.

When a classifier is configured using such an oversampling method or undersampling method, the performance of the classifier differs depending on the values of the weights N and M described above, so it is necessary to appropriately determine the weight value. Become.
In the technique described in Non-Patent Document 1, a value obtained by dividing the number of learning data of many classes by the number of learning data belonging to the minority class is set as a weight N in the oversampling method (N = the learning data belonging to the many classes). Number / number of learning data belonging to the minority class). In the technique described in Non-Patent Document 1, the number of learning data of a small number of classes and the number of learning data of a large number of classes are made the same in this way, thereby correcting the deviation in the composition ratio of the class to which the learning data belongs. .

In the technique described in Patent Document 1, the average number P of the minority class and the majority class is calculated, and a value obtained by dividing the average number P by the number of learning data belonging to the minority class is set as a weight N (N = P / the number of learning data belonging to the minority class), the inverse of the value obtained by dividing the average number P by the number of learning data belonging to the majority class, weight M (1 / M = P / number of learning data belonging to the majority class) It is said.
In the technique disclosed in Patent Literature 1, the number of learning data belonging to a minority class and the number of learning data belonging to a majority class are both matched to the average number P thereof, whereby the composition ratio of the class to which the learning data belongs. Correct the bias.

JP 2010-204966 A

Kamei Takataka, Kadota Hayato, Matsumoto Kenichi, "Application of Oversampling Method to Fault-proneness Model", Software Reliability Study Group 3rd Workshop, pp.97-103, July 2006. Kazunori Yamaguchi and Junichi Takahashi, “Basics and Mechanisms of Multivariate Analysis Understandable”, Hidekazu System, 2004

Patent Document 1 and Non-Patent Document 1 described above are based on the assumption that a classifier with high performance can be obtained by eliminating the deviation between the number of learning data belonging to the minority class and the number of learning data belonging to the majority class. .
However, this assumption is rarely true. When creating a classifier using learning data with biased class composition ratios, it is often the case that the learning data of a small number of classes is distributed among the learning data of many classes when examining the distribution of those attributes . Therefore, it is necessary to create a classifier that accurately extracts a minority class from a large number of classes.

FIG. 9 is a diagram illustrating an example of the attribute of each learning data having a biased class composition ratio.
In FIG. 9, ◯ indicates learning data for a small number of classes, and x indicates learning data for a large number of classes.
In the example shown in FIG. 9, the minority class is distributed only within the dotted circle. Therefore, the data of many classes far from the dotted circle (the data of many classes indicated by “x” located on the left side in FIG. 9) is unnecessary for learning of the classifier. Therefore, it is obviously redundant to increase the minority class learning data by oversampling so as to be equal to the number of learning data belonging to the majority class.

  On the other hand, to reduce the learning data of many classes uniformly by undersampling so that it becomes equal to the number of learning data belonging to the minority class, the learning data of many classes in the dotted circle is unnecessarily deleted. So it ’s too much. That is, the learning data of a large number of classes in the dotted circle is deleted even though the learning data is necessary to determine the boundary with the learning data of a small number of classes. .

  Based on such new knowledge, the present inventors have come to realize that the number of learning data belonging to the minority class after sampling may be equal to or less than the number of learning data belonging to the majority class. That is, it has been recognized that there are optimum values of the weights N and M between “1” and less than “the number of learning data belonging to the majority class / the number of learning data belonging to the minority class”.

  The present invention has been made on the basis of such recognition. A classifier for determining which of two classes a given data belongs to is used learning data to which the class is assigned. If the number of learning data belonging to one class is extremely smaller than the number of the other class, the classifier can sample the learning data without excess or deficiency and classify the data with high accuracy. The purpose is to build.

  The classifier creating apparatus according to the present invention uses, as a weight, the number of learning data of a minority class, which is a class to which a relatively small number of learning data belongs, among learning data of which two classes are known. By increasing at a corresponding magnification, and reducing at least one of the number of learning data of a large number of classes, which is a class to which a relatively large number of learning data belongs, by a magnification according to the inverse of the weight A classifier that creates new learning data by changing the number of the learning data, and uses the new learning data to create a classifier for judging which of the two classes the given data belongs to A creation device, wherein the number of learning data of the majority class is equal to or less than a value obtained by dividing the number of learning data of the minority class by the number of learning data of the minority class. A provisional weight determination means for determining a plurality of provisional weights having different values from the range of the upper limit value and the lower limit value, and the number of learning data of the minority class at a magnification according to the provisional weight The new learning data is obtained by performing at least one of increasing the number of learning data of the multiple classes by a factor corresponding to the reciprocal of the temporary weight for each of the plurality of temporary weights. Learning data sampling means for creating a plurality of sets of learning data, and using the new learning data after the number of learning data is changed by the learning data sampling means, the given learning data is in any of the two classes The provision of a temporary classifier for determining whether or not it belongs to each of the new learning data sets, a learning means for obtaining a plurality of temporary classifiers, and the temporary classifier Calculating the evaluation value for evaluating the performance of the temporary classifier obtained by the learning means based on the result of classifying the learning data into one of the two classes. An evaluation value calculation unit that obtains a plurality of evaluation values, an evaluation value obtained by the evaluation value calculation unit, and a temporary weight used in obtaining the evaluation value An optimum weight deriving means for deriving a weight corresponding to an evaluation value having the largest value in a range from the lower limit value to the upper limit value as an optimum value of the weight, and obtaining a relationship between the value and the weight; , Using the optimum value of the weight as the weight, increasing the number of learning data of the minority class by a factor according to the weight, and multiplying the number of learning data of the majority class by the inverse of the weight Reducing with A classification that creates at least one of the above, creates the new learning data, and uses the new learning data to create a classifier for determining which of the two classes the given learning data belongs to And a container creating means.

  The classifier creating method of the present invention uses the number of learning data of a minority class, which is a class to which a relatively small number of learning data belongs, among learning data that is known to belong to which of two classes. By increasing at a corresponding magnification, and reducing at least one of the number of learning data of a large number of classes, which is a class to which a relatively large number of learning data belongs, by a magnification according to the inverse of the weight A classifier that creates new learning data by changing the number of the learning data, and uses the new learning data to create a classifier for judging which of the two classes the given data belongs to The method is a creation method, wherein the number of learning data of the majority class is equal to or less than a value obtained by dividing the number of learning data of the minority class by the number of learning data of the minority class, and the upper limit value and A provisional weight determination step for determining a plurality of provisional weights having different values from the range of the upper limit value and the lower limit value, and the number of learning data of the minority class at a magnification according to the provisional weight The new learning data is obtained by performing at least one of increasing the number of learning data of the multiple classes by a factor corresponding to the reciprocal of the temporary weight for each of the plurality of temporary weights. The learning data sampling step for creating a plurality of sets of the learning data, and using the new learning data after the number of learning data is changed by the learning data sampling step, the given learning data is in any of the two classes Creating a temporary classifier for determining whether it belongs to each of the new learning data sets, and obtaining a plurality of temporary classifiers by the temporary classifier; Calculating the evaluation value for evaluating the performance of the temporary classifier obtained by the learning step based on the result of classifying the learning data into one of the two classes. An evaluation value calculation step for obtaining a plurality of evaluation values, an evaluation value obtained by the evaluation value calculation step, and a temporary weight used when obtaining the evaluation value are performed each time. An optimum weight deriving step of deriving a weight corresponding to an evaluation value having the largest value in a range from the lower limit value to the upper limit value as an optimum value of the weight in the obtained relationship. , Using the optimum value of the weight as the weight, increasing the number of learning data of the minority class by a factor according to the weight, and multiplying the number of learning data of the majority class by the inverse of the weight Reducing with A classification that creates at least one of the above, creates the new learning data, and uses the new learning data to create a classifier for determining which of the two classes the given learning data belongs to And a container creation step.

  The computer program according to the present invention causes a computer to function as each means of the classifier creating apparatus.

  According to the present invention, the upper limit value and the lower limit value of the weights are determined from the range of one or more values less than or equal to the number of learning data of the majority class divided by the number of learning data of the minority class. The weight corresponding to the evaluation value having the highest value as the evaluation value for the classifier is set as the optimum value of the weight from the range of the value and the lower limit value, and the learning data is sampled using the optimum value of the weight and the classifier Was created. Therefore, even if the number of learning data belonging to one class is extremely smaller than the number of learning data belonging to the other class, a learning class can be sampled without excess or deficiency and a classifier that can classify the data with high accuracy is constructed. be able to.

It is a figure which shows an example of a structure of a classifier preparation apparatus. It is a figure which shows an example of a decision tree. It is a figure explaining an example of a mixing matrix. It is a figure which shows an example of the relationship between F value and weight N. It is a flowchart explaining an example of a process of a classifier preparation apparatus. It is a flowchart explaining the detail of a learning data sampling process. It is a flowchart explaining optimal weight derivation processing. It is a figure which shows the mixing matrix in an Example and a comparative example. It is a figure which shows an example of the attribute of each learning data with which the composition ratio of a class has bias.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
[First Embodiment]
First, a first embodiment of the present invention will be described.
(Configuration of classifier creating apparatus 100)
FIG. 1 is a diagram illustrating an example of the configuration of the classifier creating apparatus 100. The classifier creating apparatus 100 shown in FIG. 1 is realized by using, for example, a computer system (information processing apparatus) including a CPU, ROM, RAM, HDD, and various interfaces. The classifier created by the classifier creating apparatus 100 applies an attribute of given data to a rule, and determines which of the two classes the data belongs to. Below, an example of the function which the classifier preparation apparatus 100 has is demonstrated. In the present embodiment, a case where a rule of a classifier is constructed by a decision tree will be described as an example. The decision tree is one of data analysis methods, and as shown in FIG. 2, is an analysis method that classifies data like branches and leaves of trees according to various conditions. In the example shown in FIG. 2, the data is finally classified into one of a good product and a defective product.

<Learning data input unit 101>
The learning data input unit 101 inputs a plurality of learning data based on an operation by an operator (creator of the classifier) or the like. Each of the plurality of learning data is performance data for which a correct class is known. The learning data is, for example, measurement data of a product by a sensor, and the class is, for example, whether or not the product is a non-defective product. In this case, each of the plurality of learning data is provided (included) with information indicating whether the product that is the basis of the learning data is a non-defective product or a defective product. About the number of learning data, it can determine suitably by the performance anticipated with respect to the classifier to produce, the calculation load at the time of creating a classifier, etc.

  The learning data input unit 101 is realized, for example, when the CPU inputs learning data based on an operation of a user interface by an operator and stores it in a RAM or the like. In addition, for example, the CPU may input learning data from an external device via a communication interface and store it in a RAM or the like. Further, the CPU may input (read out) learning data stored in a portable storage medium and store it in a RAM or the like.

<Weight range input unit 102>
The weight range input unit 102 inputs a weight N range based on an operation by an operator or the like. As described above, the present inventors have proposed that a new optimum weight value exists between “1” and “the number of learning data belonging to the majority class / the number of learning data belonging to the minority class”. Obtained knowledge. Therefore, the lower limit value N _min of the range of the weight N input by the weight range input unit 102 is “1” or more (N _min ≧ 1), and the range of the weight N input by the weight range input unit 102 is The range of the weight N is determined by the operator so that the upper limit value N _max is equal to or less than “(number of learning data belonging to majority class / number of learning data belonging to minority class) N _r ” (N _max ≦ N _r ). It is done.

  The weight range input unit 102 is realized, for example, when the CPU inputs a range of the weight N based on the operation of the user interface by the operator and stores it in the RAM or the like. In addition, for example, the CPU may input the range of the weight N from an external device via the communication interface and store it in the RAM or the like. Further, the CPU may input (read out) the range of the weight N stored in the portable storage medium and store it in the RAM or the like.

<Temporary Weight Determination Unit 103>
The temporary weight determination unit 103 uses a preset method from the range of the weight N input by the weight range input unit 102 so that the preset number of temporary weights N ′ do not overlap. decide. In the present embodiment, a case where _three temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ are determined by the following formulas (1) to (3) will be described as an example.
Temporary weight N ′ ₁ = N _min (1)
Temporary weight N ′ ₂ = (N _min + N _max ) / 2 (2)
Temporary weight N ′ ₃ = N _max (3)

As described above, in this embodiment, the lower limit value N _min of the range of the weight N is “1” or more, and the upper limit value N _max is “N _r ” or less. Thus, when the N _min = 1 and N _max = N _r (1) When using the ~ (3) determining the weights of the temporary weight _N'1 The temporary "1", the temporary weight N _'2' (1 + N _r) / 2 ", and the weight _N'3 provisional is" N _r ".
The provisional weight determination method may be other than the equations (1) to (3). Since the present invention is a method for predicting the optimum value of the weight N from a plurality of provisional weights and respective evaluation values within the range of N _min to N _max , the provisional weight value selection method is respectively The provisional weight values may be selected so as to cover a wide range from N _min to N _max without overlapping. For example, the temporary weights may be determined using the following formulas (1) ′ to (3) ′.
Temporary weight N ′ ₁ = N _min + 1 × (N _max −N _min ) / 4 (1) ′
Temporary weight N ′ ₂ = N _min + 2 × (N _max −N _min ) / 4 (2) ′
Temporary weight N ′ ₃ = N _min + 3 × (N _max −N _min ) / 4 (3) ′ For example, the temporary weight determination unit 103 reads the range of the weight N from the RAM or the like. , (1) to (3) are calculated to obtain the value of the temporary weight N ′ and store it in the RAM or the like.

<Learning data sampling unit 104>
The learning data sampling unit 104 uses, for each of the plurality of temporary weights N ′ determined by the temporary weight determination unit 103, learning data belonging to the minority class among the learning data input by the learning data input unit 101. Oversampling is performed to create new learning data in which the number of learning data belonging to the minority class is increased.
The learning data sampling unit 104 obtains new learning data for each temporary weight N ′. In the present embodiment, three sets of new learning data are obtained by the learning data sampling unit 104 corresponding to the _three temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ . In the present embodiment, the learning data sampling unit 104 includes three learning data sampling units (a first learning data sampling unit 104a, a second learning data sampling unit 104b, A third learning data sampling unit 104c) is included. The first, second, and third learning data sampling units 104a, 104b, and 104c oversample the learning data belonging to the minority class using temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ , respectively.

The functions of the first, second, and third learning data sampling units 104a, 104b, and 104c are the same except that the value of the temporary weight N ′ to be used is different. Therefore, here, the second learning data sampling unit 104b that samples the learning data belonging to the minority class using the temporary weight N ′ ₂ will be described, and the other first and third learning data sampling units 104a and 104c. The detailed description about is omitted.

The second learning data sampling unit 104b always adds many classes of learning data to the new learning data. The second learning data sampling unit 104b, a learning data belonging to a few classes, on average and repeatedly many times the weight _N'second temporary replication to add to the new training data.
Specifically, the second learning data sampling unit 104b sets a data number i for each learning data to which a data number i (i = 0, 1, 2,...) For identifying learning data is assigned. The following processing is performed in order from the smallest learning data.

  First, the second learning data sampling unit 104b determines whether the learning data with the data number i is learning data belonging to the minority class. As a result of this determination, when the learning data with the data number i is not learning data belonging to the minority class (when it is learning data belonging to the majority class), the second learning data sampling unit 104b uses the learning data with the data number i. Adopt as new learning data.

On the other hand, when the learning data with the data number i is learning data belonging to the minority class, the second learning data sampling unit 104b accumulates the additional number of learning data p (p = 0, 1, 2,...). Until the value becomes equal to or greater than the upper limit q (the initial value of q is a temporary weight N ′ ₂ ), the learning data with the data number i is repeatedly adopted as new learning data. The second learning data sampling unit 104b a value obtained by adding the temporary weight _N'2 to the current upper limit value q as a new upper limit value q.
For example, when the value of the provisional weight N ′ ₂ is “2.5”, the learning data belonging to the minority class is used as new learning data by repeating it alternately three times and twice in order from the smallest data number i. Is done.
The learning data sampling unit 104 creates new learning data by oversampling the learning data belonging to the minority class based on each of the temporary weights N ′, for example, by the CPU performing the above processing. This is realized by storing learning data in a RAM or the like.

<Learning unit 105>
The learning unit 105 learns the new learning data set obtained by the learning data sampling unit 104 for each new learning data set, and the correct answer rate (ratio in which the class predicted by the classifier matches the actual class) As many temporary classifiers as possible are created as many as the number of new learning data sets.
In the present embodiment, the learning data sampling unit 104 obtains three sets of new learning data. Therefore, three temporary classifiers are obtained by the learning unit 105. The learning unit 105 includes three learning units (a first learning unit 105a, a second learning unit 105b, and a third learning unit 105c) in order to individually create these three temporary classifiers. The first, second, and third learning units 105a, 105b, and 105c learn temporary learning data obtained by the first, second, and third learning data sampling units 104a, 104b, and 104c, respectively. Create a classifier for.

A learning algorithm for creating a decision tree that is a classifier can be realized by using a known technique described in Non-Patent Document 2, for example. Therefore, detailed description of the learning algorithm for creating the classifier is omitted here.
In the learning unit 105, for example, the CPU learns new learning data for each set of new learning data, creates the same number of temporary classifiers as the number of new learning data sets, and information on the created temporary classifiers Is stored in a RAM or the like.

<Evaluation Value Calculation Unit 106>
The evaluation value calculation unit 106 evaluates each of the plurality of temporary classifiers obtained by the learning unit 105 using the original learning data, and calculates an evaluation value. In the present embodiment, since three temporary classifiers are created by the learning unit 105, three evaluation values are obtained. In order to obtain these three evaluation values, the evaluation value calculation unit 106 includes three evaluation value calculation units (a first evaluation value calculation unit 106a, a second evaluation value calculation unit 106b, and a third evaluation value calculation unit 106c. ). The first, second, and third evaluation value calculation units 106a, 106b, and 106c calculate the evaluation values of the classifiers obtained by the first, second, and third learning units 105a, 105b, and 105c, respectively. .

  In the present embodiment, a case where the F value described in Non-Patent Document 1 or the like is used as an evaluation value of the classifier will be described as an example. The functions of the first, second, and third evaluation value calculation units 106a, 106b, and 106c are the same except that the classifiers to be evaluated are different. Therefore, here, the first evaluation value calculation unit 106a will be described, and detailed descriptions of the other second and third evaluation value calculation units 106b and 106c will be omitted.

First, the first evaluation value calculation unit 106a inputs the original learning data input by the learning data input unit 101 into the temporary classifier created by the first learning unit 105a and classifies the mixed learning matrix. Create
FIG. 3 is a diagram illustrating an example of the mixing matrix. In FIG. 3, n ₁₁ is the number of correctly predicted learning data having a large number of actual results (actual classes), and n ₁₂ is the number of erroneously predicted learning data having a large number of actual results. Also, n ₂₂ is the number of correctly predicted learning data with a small number of classes, and n ₂₁ is the number of incorrectly predicted learning data with a small number of classes.

Using these values n ₁₁ , n ₁₂ , n ₂₁ , and n ₂₂ , the first evaluation value calculation unit 106a uses the precision shown in the following formula (4) and the recall shown in the following formula (5). Calculate
Precision = n ₂₂ / (n ₁₂ + n ₂₂ ) (4)
Reproducibility = n ₂₂ / (n ₂₁ + n ₂₂ ) (5)
The relevance ratio is a ratio of learning data that actually belongs to the minority class among the learning data predicted to be a minority class. For example, if the learning data is sensor measurement data of a sensor and the class is whether the product is good or not, this conformance rate is calculated based on whether the product is actually defective or not. It indicates how much the product is included.
The recall is the ratio of learning data that is predicted to be a minority class among the learning data that actually belongs to the minority class. In the example described above, this recall rate indicates how much a product that is predicted to be defective is included in a really defective product.
Both the recall ratio and the matching ratio indicate that the larger the value, the higher the evaluation.

  Next, the first evaluation value calculation unit 106a uses the relevance ratio and the recall ratio, and the adjustment coefficient β (usually 1), which is set in advance, as shown in the following formula (6). Calculate the F value.

As shown in the equation (6), the F value is a weighted harmonic average value of the precision and the recall.
As described above, the first, second, and third evaluation value calculation units 106a, 106b, and 106c evaluate the classifiers obtained by the first, second, and third learning units 105a, 105b, and 105c. F values F ₁ , F ₂ and F ₃ are obtained as values.
The evaluation value calculation unit 106 is realized, for example, when the CPU calculates an evaluation value (F value) for each of a plurality of temporary classifiers and stores it in a RAM or the like.

<Optimum Weight Deriving Unit 107>
The optimum weight deriving unit 107 uses the plurality of temporary weights N ′ obtained by the temporary weight determining unit 103 and the plurality of evaluation values F obtained by the evaluation value calculating unit 106 corresponding to the weights. The optimum value N _opt is calculated. In this embodiment, three weight provisional _{N'1, N'2,} and _N'3, by using the three F values F _1, F _2, F ₃ corresponding to those of the weighting optimal value N _opt Is obtained.
In the present embodiment, the relationship F (N) between the weight N and the F value is approximated by a quadratic function expressed by the following equation (7). The optimum weight deriving unit 107 obtains the following (7) from the combinations (N ′ ₁ , F ₁ ), (N ′ ₂ , F ₂ ), (N ′ ₃ , F ₃ ) of the _three temporary weights and the F value. ) The coefficients a, b, and c of the quadratic function shown in the equation are calculated by the following equations (8) to (10).

In the equation (7), F (N) indicates that the F value is represented by a function of N, and N indicates a weight variable (weight).
The optimum weight deriving unit 107 substitutes the coefficients a, b, c calculated by the equations (8) to (10) and the lower limit value “N _min ” of the weight N range into the equation (7). F value F (N _min ) when the value of weight N is N _min is calculated. Also, the optimum weight deriving unit 107 substitutes the coefficients a, b, and c calculated by the equations (8) to (10) and “N _max ” that is the upper limit value of the weight N range into the equation (7). Then, the F value F (N _max ) when the value of the weight N is N _max is calculated.

Based on the coefficients a, b, and c calculated as described above, the optimum weight deriving unit 107 determines whether or not the quadratic function shown in the equation (7) is an upward convex function. Specifically, when the relationship of the following expression (11) is satisfied, the quadratic function shown in the expression (7) becomes a convex function.
a + b + c <0 (11)

Here, when the quadratic function passing through three different points can be expressed as shown in formula (7), and the following formula (11) is satisfied, the quadratic function shown in formula (7) is convex upward Explain that it becomes a function.
First, a quadratic function passing through any three different points (x ₁ , y ₁ ), (x ₂ , y ₂ ), (x ₃ , y ₃ ) can be expressed by the following ( 12)

Substituting the point (x ₁ , y ₁ ) into the equation (12) yields the following equation (13). Further, when the point (x ₂ , y ₂ ) and the point (x ₃ , y ₃ ) are respectively substituted into the equation (12), the following equations (14) and (15) are obtained.

From the above, the quadratic functions passing through _three different points (x ₁ , y ₁ ), (x ₂ , y ₂ ), and (x ₃ , y ₃ ) are expressed by equations (13), (14), and (15). This is a function of equation (12) having coefficients a, b, and c calculated by the equation. Here, if x ₁ is replaced by N ′ ₁ , y ₁ is replaced by F ₁ , x ₂ is replaced by N ′ ₂ , y ₂ is replaced by F ₂ , x ₃ is replaced by N ′ ₃ , and y ₃ is replaced by F ₃ , equation (12) is obtained. In Equation (7), Equation (13) corresponds to Equation (8), Equation (14) corresponds to Equation (9), and Equation (15) corresponds to Equation (10). Thus, a quadratic function passing through three different points can be expressed as in equation (7).

Next, when the equation (12) is differentiated by x, the following equation (16) is obtained. Accordingly, the extreme value x _opt of the quadratic function y is the value of x when the value on the right side of the equation (16) is 0 (zero), and is obtained as in the following equation (17).

Conditions in extreme x _opt 2 quadratic function y is convex upward, the (12) second floor the value of the differentiation is that negative at extreme x _opt of formula below (18), (19 )

Therefore, as described above, when the relationship of the expression (11) is satisfied, the quadratic function shown in the expression (7) becomes a convex function.
As described above, when the relationship of the expression (11) is satisfied (when the quadratic function shown in the expression (7) is an upward convex function), the optimum weight deriving unit 107 is expressed by the following expression (20) ( 7) The weight N _opt corresponding to the maximum value of the quadratic function shown in the equation is calculated.

The optimum weight deriving unit 107 calculates the weight N value calculated by the equation (20) to be not less than “N _min ” which is the lower limit value of the weight N range and not more than “N _max ” which is the upper limit value of the weight N range. It is determined whether it is within the range.
As a result of this determination, when the value of the weight N calculated by the equation (20) is within the range of “N _min ” or more and “N _max ” or less, the optimum weight derivation unit 107 calculates the equation by the equation (20). The weight N is set to the optimum weight value _Nopt .
On the other hand, when the value of the weight N calculated by the equation (20) is not within the range of “N _min ” or more and “N _max ” or less, the optimum weight derivation unit 107 sets “N” which is the lower limit value of the weight N range. _Among F values F (N _min ) corresponding to “ _min ” and F values F (N _max ) corresponding to “N _max ” which is the upper limit value of the weight N range, corresponding to the F value having the larger value The weight N to be used (“N _min ” or “N _max ”) is selected as the optimum weight value N _opt .

Similarly, when the relationship of the expression (11) is not satisfied (when the quadratic function shown in the expression (7) is not a convex function), the optimum weight deriving unit 107 similarly performs F (N _min ) and F ( N _max ), the weight N (“N _min ” or “N _max ”) corresponding to the larger F value is selected as the optimum weight value N _opt .

FIG. 4 is a diagram illustrating an example of the relationship between the F value and the weight N. FIG. 4 shows an example in which the quadratic function shown in the equation (7) is an upward convex function. Since the evaluation of the classifier is higher as the F value is larger, in the example shown in FIG. 4, the weight N when the F value becomes the maximum value is the optimum weight value N _opt .
The optimum weight deriving unit 107 is realized, for example, when the CPU calculates the optimum weight value _Nopt and stores it in the RAM or the like.

<Learning data sampling unit 108>
The learning data sampling unit 108 oversamples the learning data belonging to the minority class among the learning data input by the learning data input unit 101 using the optimum weight value N _opt obtained by the optimum weight deriving unit 107. Thus, new learning data in which the learning data belonging to the minority class is increased is created. The method for oversampling the learning data belonging to the minority class is realized by the same method as the method shown in the description of the learning data sampling unit 104, and thus detailed description thereof is omitted here.
In the learning data sampling unit 108, for example, the CPU oversamples the learning data belonging to the minority class based on the optimum weight value _Nopt to create new learning data, and stores the new learning data in the RAM or the like. It is realized by.

<Learning unit 109, classifier storage unit 110>
The learning unit 109 learns the new learning data obtained by the learning data sampling unit 108, creates (one) a classifier having the highest accuracy rate, and stores it in the classifier storage unit 110. As described above, a learning algorithm for creating a decision tree that is a classifier can be realized using a known technique. Therefore, here, a detailed description of the learning algorithm for creating a classifier is given. Omitted.
The learning unit 109 is realized, for example, when the CPU learns new learning data to create a classifier and stores information on the created classifier in an HDD or the like. The classifier storage unit 110 is realized by an HDD or the like, for example.

  The classifier creating apparatus 100 uses the classifier obtained as described above to determine which of the two classes data to be evaluated belongs, and outputs the determined result. Examples of output forms include display on a display device, storage on a storage medium, transmission to an external device, and the like. In this way, even if the classifier creating apparatus 100 creates both a classifier and classifies data using the classifier, the classifier created by the classifier creating apparatus 100 is used. Alternatively, the information may be transferred to another information processing apparatus different from the classifier creating apparatus 100, and the information processing apparatus may perform data classification using the classifier.

(Operation flowchart)
Next, an example of processing of the classifier creating apparatus 100 will be described with reference to the flowchart of FIG.
First, in step S501, the learning data input unit 101 inputs a plurality of learning data based on an operation by an operator or the like.
Next, in step S502, the weight range input unit 102 inputs a range of weight N. The weight range input unit 102 only when the lower limit value N _min of the input weight N range is “1” or more and the upper limit value N _max of the weight N range is “N _r ” or less. The input weight N range is received. In other cases, the operator is prompted to input the weight N range again using a display screen or the like.

Next, in step S503, the temporary weight determining unit 103, from the range of the input weighting N in step S502, the weight _N'1 three temporary, _N'2, determines the _N'3.
Next, in step S504, the learning data sampling unit 104 selects the temporary weights N ′ ₁ , N ′ ₂ and N ′ ₃ determined in step S503 one by one in this order.
Next, in step S505, the learning data sampling unit 104 performs learning data sampling processing for oversampling the learning data belonging to the minority class using the temporary weight N ′ selected in step S504, and creating new learning data. Do. The details of the learning data sampling process will be described later with reference to FIG.

Next, in step S506, the learning unit 105 learns the new learning data obtained in step S505, and creates a temporary classifier that maximizes the correct answer rate.
Next, in step S507, the evaluation value calculation unit 106 calculates an F value, which is an evaluation value for the temporary classifier obtained in step S506, using the learning data input in step S501.

Next, in step S508, the learning data sampling unit 104 determines whether or not all of the temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ determined in step S503 (ie, N ′ ₃ ) have been selected. To do. As a result of this determination, if all of the temporary weights N ′ ₁ , N ′ ₂ , N ′ ₃ determined in step S503 have not been selected, the process returns to step S504. Then, for all of the temporary weights N ′ ₁ , N ′ ₂ , N ′ ₃ determined in step S503, the steps until the generation of the temporary classifier and the derivation of the evaluation value (F value) are completed. The processes of S504 to S508 are repeated.

If it is determined in step S508 that all the temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ determined in step S503 have been selected, the process proceeds to step S509.
In step S509, the optimum weight deriving unit 107 determines the temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ obtained in step S503, and the plurality of values obtained by the evaluation value calculation unit 106 corresponding to them. The optimum weight derivation process for deriving the optimum weight value N _opt is performed using the F values F ₁ , F ₂ , and F ₃ . The details of the optimum weight derivation process will be described later with reference to FIG.

Then, over at step S510, the learning data sampling unit 108 performs the same training data sampling processing in step S505, the learning data which belongs to a few classes using the optimum value N _opt of the weights obtained in step S509 Sampling and creating new learning data.
Next, in step S511, the learning unit 109 learns the new learning data obtained in step S510, creates (one) classifier with the highest accuracy rate, and stores it in the classifier storage unit 110. Store.
And the process by the flowchart of FIG. 5 is complete | finished.

Next, details of the learning data sampling process in steps S505 and S510 in FIG. 5 will be described with reference to the flowchart in FIG. In the learning data sampling process of step S505 in FIG. 5, the following processes are performed in the order of the first, second, and third learning data sampling units 104a, 104b, and 104c. Also, the learning data sampling processing in step S510 in FIG. 5, that the temporary weight N'the optimum value N _opt weight is only different from the learning data sampling processing in step S505. Therefore, only the details of the learning data sampling process in step S510 of FIG. 5 will be described here.

First, in step S601, the learning data sampling unit 104 sets the value of the data number i to an initial value (= 0 (zero)), the value of the accumulated additional number p of learning data to an initial value (= 0 (zero)), and learning. The upper limit value q of the cumulative additional number p of data is set as an initial value (= temporary weight N ′ selected in step S504 in FIG. 5).
Next, in step S602, the learning data sampling unit 104 selects learning data with the data number i from the learning data input in step S501 of FIG.

Next, in step S603, the learning data sampling unit 104 determines whether the learning data with the data number i is learning data belonging to the minority class.
As a result of the determination, if the learning data with the data number i is learning data belonging to the minority class, the process proceeds to step S607 described later. On the other hand, when the learning data with the data number i is not learning data belonging to the minority class (when it is the majority class), the process proceeds to step S604.

In step S604, the learning data sampling unit 104 adopts the learning data with the data number i as new learning data.
Next, in step S605, the learning data sampling unit 104 determines whether all of the learning data input in step S501 of FIG. 5 has been selected. If all of the learning data is selected as a result of this determination, the processing according to the flowchart of FIG. 6 ends.
On the other hand, if not all of the learning data has been selected, the process proceeds to step S606.

In step S606, the learning data sampling unit 104 adds “1” to the data number i and updates the data number i.
And the process after step S602 is performed with respect to the learning data of the following data number i.
As described above, if the result of determination in step S603 is that the learning data with data number i is learning data belonging to the minority class, the process proceeds to step S607.

In step S607, the learning data sampling unit 104 determines whether or not the accumulated additional number p of learning data is less than the upper limit q (p <q). As a result of the determination, if the value of the cumulative additional number p of learning data is less than the upper limit q, the process proceeds to step S608.
In step S608, the learning data sampling unit 104 employs the learning data with the data number i as new learning data.
Next, in step S609, the learning data sampling unit 104 adds “1” to the accumulated additional number p of learning data, and updates the value of the accumulated additional number p of learning data.

Then, the process proceeds to step S607, and the processes of steps S607 to S609 are repeated until the value of the cumulative additional number p of learning data becomes equal to or greater than the upper limit q.
When the value of the cumulative additional number p of learning data becomes equal to or greater than the upper limit q (p ≧ q), the process proceeds to step S610. In step S610, the learning data sampling unit 104 adds the provisional weight N ′ selected in step S504 in FIG. 5 to the upper limit q of the accumulated addition number p of learning data, thereby accumulating the addition of learning data. The upper limit value q of the number p is updated. Then, the process proceeds to step S605 described above.
Through the processes in steps S607 to S610 described above, learning data belonging to the minority class is repeatedly adopted as new learning data on the average by the number of temporary weights N ′.

Next, the optimum weight derivation process in step S509 in FIG. 5 will be described with reference to the flowchart in FIG.
First, in step S701, the optimal weight derivation unit 107, three sets of weights and F values of the provisional _{(N'1, F 1),} (N'2, F 2), (N'3, F 3) From the above, the coefficients a, b, and c of the quadratic function shown in Equation (7) are calculated (see Equations (8) to (10) for the calculation equations).
Next, in step S702, the optimum weight deriving unit 107 substitutes the coefficients a, b, and c obtained in step S701 and “N _min ” that is the lower limit value of the weight N range into the equation (7). Thus, the F value F (N _min ) when the value of the weight N is N _min is calculated. Further, the optimum weight deriving unit 107 substitutes the coefficients a, b, and c obtained in step S701 and “N _max ”, which is the upper limit value of the weight N range, in the equation (7), and The F value F (N _max ) when the value is N _max is calculated.

Next, in step S703, the optimum weight deriving unit 107 determines whether or not the quadratic function shown in the equation (7) is an upward convex function based on the coefficients a, b, and c obtained in step S701. Determine. When the coefficients a, b, and c obtained in step S701 satisfy the relationship of equation (11), it is determined that the quadratic function shown in equation (7) is an upward convex function.
If the result of this determination is that the quadratic function shown in equation (7) is a convex function, the process proceeds to step S704.
In step S704, the optimum weight deriving unit 107 calculates a weight N corresponding to the maximum value of the quadratic function shown in equation (7) (see equation (20)).

Next, in step S705, the optimum weight deriving unit 107 has the weight N value obtained in step S704 equal to or higher than the lower limit value “N _min ” of the weight N range and the upper limit value of the weight N range. It is determined whether or not it is within the range of “N _max ”.
As a result of the determination, if the value of the weight N obtained in step S704 is in the range of “N _min ” or more and “N _max ” or less, the process proceeds to step S706.
In step S706, the optimum weight deriving unit 107 employs the weight N obtained in step S704 as the optimum weight value _Nopt . And the process by the flowchart of FIG. 7 is complete | finished.

If it is determined in step S703 that the quadratic function shown in equation (7) is not an upward convex function, and in step S705, the value of the weight N obtained in step S704 is equal to or greater than “N _min ”. If it is determined that it is not within the range of “N _max ” or less, the process proceeds to step S707.
In step S707, the optimum weight deriving unit 107 determines that the F value F (N _min ) corresponding to the lower limit value “N _min ” of the weight N range is the upper limit value “N _max ” of the weight N range. It is determined whether or not it is less than the F value F (N _max ) corresponding to “”.
If the condition is satisfied as a result of the determination, the process proceeds to step S708.

In step S708, the optimum weight deriving unit 107 employs “N _max ” that is the upper limit value of the weight N range as the optimum weight value N _opt . And the process by the flowchart of FIG. 7 is complete | finished.
On the other hand, if the determination condition in step 707 is not satisfied, the process proceeds to step S709. In step S709, the optimum weight deriving unit 107 employs “N _min ”, which is the lower limit value of the weight N range, as the optimum weight value N _opt . And the process by the flowchart of FIG. 7 is complete | finished.

(Example)
Next, examples of the present invention will be described. In the present embodiment, a case will be described in which a classifier for determining whether or not the steel product passes the generation process in the manufacturing process from data obtained after the steel product is manufactured or in the manufacturing process is described. .
In the process of manufacturing steel products, workers inspect the rolled and cooled steel products. And if the bend of the steel product is large, the bend is corrected in the straightening process, or if the surface of the steel product is wrinkled, the wrinkle is polished with a grinder in the maintenance process, etc. After that, there is a process for determining the presence or absence of passage. Such a process is called a generation process. When the product passes the generation process, the manufacturing period is extended. For this reason, when the ordered product is manufactured, it is necessary to predict whether or not the generation process has passed before manufacturing and start manufacturing in time for the delivery date of the customer. However, whether or not the generation process passes is not uniquely determined from the product specifications (size, hardness, etc.). For this reason, the presence or absence of the passage of the generation process is predicted from past operation performance data, and the production start date is determined based on the prediction.

  As a comparative example for this example, a decision tree that predicts whether or not a generation process has passed in a certain steel manufacturing process was created without weighting (corresponding to weight = 1). As a decision tree learning algorithm, a method based on an information entropy gain ratio called C5.0 described in Non-Patent Document 2 was used. FIG. 8A is a confusion matrix representing the performance of this decision tree. As shown in FIG. 8 (a), the actual number of steel products predicted to pass this generation process was 1467, although 51,510 steel products passed the generation process. The F value was 2.44 [%].

Using the same decision tree learning algorithm, a decision tree was created by the method described in this embodiment. Here, the value of the adjustment coefficient β when calculating the F value is set to “1” (β = 1). Further, the lower limit value N _min of the weight N range is set to “1”, and the upper limit value N _max of the weight N range is set to “10”. The number of temporary weights N ′ is “3”. Further, provisional weights N ′ ₁ , N ′ ₂ , and N ′ ₃ were obtained by the equations (1), (2), and (3). That is, the value of the temporary weight N ′ ₁ is set to “1” (N ′ ₁ = 1), the value of the temporary weight N ′ ₂ is set to “5.5” (N ′ ₂ = 5.5), The value of the weight N ′ ₃ was set to “10” (N ′ ₃ = 10).

Using these temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ , the learning data belonging to the minority class was oversampled by the method described in the present embodiment to obtain three sets of new learning data. Then, these three new learning data sets were applied to the above-described decision tree learning algorithm to create three decision trees (provisional classifiers). F values F ₁ , F ₂ , and F ₃ , which are evaluation values of the three temporary classifiers created using the temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ , are 2.44 [%] (= 0.0244), 28.9 [%] (= 0.289), 25.4 [%] (= 0.254) (F ₁ = 0.0244, F ₂ = 0.289, F ₃ = 0.254).

Then, the relationship between the weight N and the F value (F (N)) was approximated by a quadratic function shown in the equation (7) to obtain the optimum weight value _Nopt . As a result, the value of the optimum value N _opt weight became 7.2243. Using this weight optimum value N _opt , new learning data was obtained by oversampling the learning data belonging to the minority class by the method described in the present embodiment. These new learning data were applied to the decision tree learning algorithm described above to create a decision tree (classifier). As a result of calculating a confusion matrix representing the performance of the decision tree (classifier) obtained in this manner, the confusion matrix is as shown in FIG. In this example, it is predicted that 99533 steel products pass the generation process, of which 22561 steel products actually pass the generation process. For this reason, the result of the present example was clearly superior to the method (comparative example) in which weighting is not performed as shown in FIG. In the comparative example shown in FIG. 8A, the F value was 2.44 [%], whereas in this example, the F value was improved to 29.9 [%].

Further, “N _r ” which is the upper limit value of the range of the weight N is represented by “the number of learning data belonging to the majority class / the number of learning data belonging to the minority class”. 7 (= 51510/656233). In other words, the optimum weight value N _opt (= 7.2243) is “1” that is the upper limit value of the range of weight N and “N _r (= 12.7)” that is the upper limit value of the range of weight N. Between. Therefore, it can be seen that in order to obtain the highest F value, it is necessary to sample so that the number of learning data belonging to the minority class is smaller than the number of learning data belonging to the majority class.

(Summary)
As described above, in the present embodiment, three temporary weights are selected from the range of the weight N designated with the lower limit value being “N _min (≧ 1)” and the upper limit value being “N _max (≦ N _r )”. N ′ ₁ , N ′ ₂ and N ′ ₃ are determined. Three sets of new learning data are created by oversampling the learning data belonging to the minority class using each of these temporary weights N ′ ₁ , N ′ ₂ , and N ′ ₃ . Three temporary classifiers are created using the three sets of new learning data that have been created, and F values F ₁ , F ₂ , and F ₃ that are evaluation values thereof are obtained. 2 representing the relationship between the weight N and the F value from the pair (N ′ ₁ , F ₁ ), (N ′ ₂ , F ₂ ), and (N ′ ₃ , F ₃ ) of the _three temporary weights and the F value. The coefficients a, b, and c of the quadratic function are calculated, the weight N corresponding to the maximum value of the quadratic function is obtained, and when the weight N is “N _min ” or more and “N _max ” or less, The weight N is set to the optimum value N _opt for the weight, and a classifier is constructed using new learning data created based on the optimum value N _opt for the weight. On the other hand, when the weight N is not “N _min ” or more and “N _max ” or less, the optimal value N _{opt of the} weight corresponding to the larger one of the corresponding F values among “N _min ” and “N _max ”. And a classifier is constructed using the new learning data created based on the optimum weight value _Nopt .
That is, in order to create a classifier capable of predicting as accurately as possible, the learning data belonging to the minority class is sampled using the optimum value N _{opt of the} weight with the highest F value. Therefore, even when the number of learning data belonging to one class is extremely smaller than the number of learning data belonging to the other class, it is possible to construct a highly accurate classifier by sampling the learning data without excess or deficiency.
In this embodiment, the relationship between the weight N and the F value is approximated to be a quadratic function, and the optimum weight value N _opt is calculated from a plurality of temporary weights N ′ and the corresponding F values. Was calculated. By doing so, it is not necessary to perform the convergence calculation, so that the calculation load when calculating the optimum weight value _Nopt can be reduced as much as possible.

(Modification)
In the present embodiment, a range of weight N is input, and a plurality of provisional weights N ′ are determined to be equally spaced within the range. However, this is not always necessary. For example, the values of a plurality of temporary weights N ′ may be directly input based on an operator's operation or the like. However, in this case, it is necessary to accept only values of “N _min ” or more and “N _max ” or less as the values of the plurality of temporary weights N ′.

Further, the lower limit value N _min of the range of the weight N is not limited to “1”. For example, if the existence range of the approximate value of the optimum weight N _opt is known from past experience, a value exceeding “1” may be adopted as the lower limit value N _min of the weight N range. Further, the upper limit value N _max of the range of the weight N is not limited to “(number of learning data belonging to the majority class / number of learning data belonging to the minority class) N _r ”. For example, when the existence range of the approximate value of the appropriate weight N is known from past experience, a value lower than “N _r ” may be employed as the upper limit value N _max of the weight N range. Further, when the value of “N _r ” is larger than a predetermined value (for example, 10), a predetermined value (for example, 10) may be adopted as the upper limit value N _max of the weight N range.
As described above, the necessity of setting the range of the weight N (N _{min to} N _max ) to other than 1 to N _r will be described. Since the present invention is a technique for predicting the optimum weight N _opt by approximating the relationship between the weight N, which is a continuous function, and the evaluation value by a plurality of temporary weights and the evaluation value, the optimum weight N _opt If the existence range is known from past experience, who range in weight N is set to present a range of the optimal weight N _opt is, the higher the prediction accuracy of optimal weight N _opt, high performance classification It is possible to create a container. Further, when the bias of the learning data is large, _Nr becomes excessive, and if the small number data is oversampled with the weight _Nr , the number of new learning data becomes enormous, and the storage device (RAM, etc.) of the computer system Even if the classifier cannot be calculated or stored in the storage device, the classifier calculation time may be enormous. In such a case, the upper limit value N _max of the range of the weight N is set to a value smaller than “N _r ” according to the learning algorithm for creating the classifier and the storage capacity of the computer system. Calculation problems can be prevented.

  Further, the method of oversampling the learning data belonging to the minority class for each temporary weight N ′ is not limited to the method described above. For example, when the value of the temporary weight N ′ has an integer part and a decimal part (for example, “2.5”), new learning data can be determined as follows. First, all of the original learning data is adopted as new learning data. Next, the learning data belonging to the minority class is adopted (repeatedly) as new learning data by the number of times that 1 is subtracted from the value of the integer part (for example, once (= 2-1)). Finally, the number of learning data belonging to the minority class is multiplied by a value after the decimal point (for example, the number of learning data belonging to the minority class is multiplied by 0.5), using random numbers or the like Thus, the learning data belonging to the minority class is randomly selected so that the values do not overlap, and the learning data belonging to the selected minority class is adopted as new learning data. Even in this way, the learning data belonging to the minority class can be sampled.

  Further, the evaluation value for the temporary classifier is not limited to the F value. For example, a weighted average value of the precision and the recall may be adopted as the evaluation value using the precision and the recall. In addition, a G average, which is a geometric mean of the TN rate expressed by the following formula (21) and the TP rate (same as the above-described recall rate), is calculated as an evaluation value by the following formula (22). May be.

Further, the number of provisional weights N ′ may be any number as long as it is plural. When the number of provisional weights N ′ is 4 or more, the relation F (N) between the weights N and the F values is approximated by a quadratic function expressed by the following equation (23). F value (F ( _{N ′ k} )) when the temporary weight N ′ _k is substituted for N, and the F value (F _k ) calculated by the evaluation value calculation unit 106 for the temporary weight _{N ′} The coefficients a, b, and c are calculated using the least square method so that the value of the sum of squares J of the errors is minimized (see equation (24)), and the calculated coefficients a, b, and c are used. As described above, the optimum weight value N _opt can be obtained. If the value of the provisional weight N ′ is 4 or more, the relationship between the weight N and the F value (F (N)) is not a quadratic function but an n-order function (n is an integer of 3 or more). , The maximum value between “N _min ” which is the lower limit value of the weight N range and “N _max ” which is the upper limit value of the weight N range is obtained as the optimum weight value N _opt. it can. Further, the relationship F (N) between the weight N and the F value may be a linear function. In this case, the optimum value N _opt of the weight can be obtained by performing the processing after NO in step S703 of FIG. However, it is desirable that the relationship F (N) between the weight N and the F value is an n-order function (n is an integer of 2 or more). This is because the relationship F (N) between the weight N and the F value is generally not a linear expression.

  In the present embodiment, the case where the classifier is a decision tree has been described as an example. However, the classifier is not limited to a decision tree. For example, SVM (Support Vector Machine), neural network, linear discrimination, etc. may be used as the classifier.

[Second Embodiment]
Next, a second embodiment of the present invention will be described. In the first embodiment described above, the case where the number of learning data belonging to the minority class is increased by oversampling the learning data belonging to the minority class has been described as an example. On the other hand, in this embodiment, a case where learning data belonging to a large number of classes is undersampled to reduce the number of learning data belonging to a large number of classes will be described as an example. As described above, the present embodiment and the first embodiment mainly differ in part of the learning data sampling process. Therefore, in the description of the present embodiment, the same parts as those in the first embodiment are denoted by the same reference numerals as those in FIGS.

In the present embodiment, the learning data sampling process is realized by setting “minority class” to “majority class” and “N ′” to “1 / N ′” in the flowchart of FIG. In this case, the learning data belonging to the minority class is added to the new learning data as it is (see steps S603 and S604). On the other hand, learning data belonging to a large number of classes is adopted as new learning data on average 1 / N ′ times.
In this way, the learning data belonging to the majority class is necessarily undersampled so that the number of learning data belonging to the majority class is equal to the number of learning data belonging to the minority class, and belongs to the majority class. Of the learning data, it is possible to suppress as much as possible that the learning data that is important for creating a highly accurate classifier is not included in the new learning data. Therefore, even when the number of learning data belonging to one class is extremely smaller than the number of learning data belonging to the other class, it is possible to construct a highly accurate classifier by sampling the learning data without excess or deficiency.

  Also in the present embodiment, various modifications described in the first embodiment can be employed. When learning data belonging to a large number of classes is selected at random, learning data that is 1 / N 'times the number of learning data belonging to the large number of classes is obtained from learning data belonging to the large number of classes using random numbers or the like. , It may be selected randomly so that the values do not overlap.

[Third Embodiment]
Next, a third embodiment will be described. In the above-described first embodiment, the case of oversampling has been described as an example, and in the above-described second embodiment, the case of undersampling has been described as an example. On the other hand, in this embodiment, a case where both oversampling and undersampling are performed will be described as an example. Thus, the present embodiment and the first and second embodiments mainly differ in part of the learning data sampling process. Therefore, in the description of the present embodiment, the same parts as those in the first and second embodiments are denoted by the same reference numerals as those in FIGS.

In the present embodiment, a value of priority r representing a ratio of giving priority to an increase in learning data belonging to the minority class with respect to a decrease in the number of learning data belonging to the majority class is set based on an operation by an operator, for example. . The value of the priority r is a value in the range of 0 to 1 (0 ≦ r ≦ 1).
A provisional weight _N ′ ₊ (referred to as “temporary weight _N ′ ₊ for the minority class” in the following description) representing the increase rate of learning data belonging to the minority class is defined as the following equation (25): A temporary weight N ′ ₋ (referred to as “temporary weight N ′ ₋ for the multi-class” in the following description) representing the reduction rate of the learning data belonging to the multi-class is defined as the following equation (26).
_{N ′ +} = r × N ′ + (1-r) (25)
_{N ′ −} = _{N ′} / _{N ′ +} (26)

In Expressions (25) and (26), N ′ is the temporary weight described in the first and second embodiments.
Then, for example, in the learning data sampling process (see FIG. 6) described in the first embodiment, “temporary weight _N ′ ₊ for the minority class” is used instead of “ _N ′”, and the minority class is used. Oversample the training data. Thereafter, in the learning data sampling process (see FIG. 6) described in the second embodiment, learning data after oversampling using “1 / N ′ ₋ ” instead of “1 / N ′” Undersampling the learning data of many classes included in. The learning data obtained in this way becomes new learning data. Note that the provisional weight _N ′ ₊ for the minority class and the provisional weight _N ′ ₋ for the majority class can be obtained by the number of provisional weights N ′ (“3” in the first embodiment). Of course.

When the provisional weight _N ′ ₊ for the minority class and the provisional weight _N ′ ₋ for the majority class are determined as in the expressions (25) and (26), when the value of the priority r is “1” The temporary weight _N ′ ₊ for the minority class becomes equal to the temporary weight _N ′ shown in the first embodiment ( _{N ′ +} = _{N ′} ), and the temporary weight N ′ ₋ for the majority class The value is “1” (N ′ ₋ = 1). Therefore, when the value of the priority r is “1”, only oversampling is performed to create new learning data (that is, the same processing as in the first embodiment is performed).
On the other hand, when the value of the priority r is “0”, the temporary weight _N ′ ₊ for the minority class becomes “1” (N ′ ₋ = 1) and the temporary weight N ′ ₋ for the majority class. Is equal to the temporary weight N ′ shown in the second embodiment (N ′ ₋ = N ′). Therefore, when the value of the priority r is “0”, only undersampling is performed to create new learning data (that is, the same processing as in the second embodiment is performed).

Also, the number of learning data belonging to the minority class and the number of learning data belonging to the majority class of the new learning data are as shown in the following equations (27) and (28), respectively.
Number of learning data belonging to a minority class of new learning data = Number of learning data belonging to a minority class of original learning data × _{N ′ +} (27)
Number of learning data belonging to many classes of new learning data = Number of learning data belonging to many classes of original learning data × _{N ′ +} / _{N ′} (28)

Therefore, the value of the temporary weight N ′ is “N _r (= the number of learning data belonging to the majority class of the original learning data / the learning data belonging to the minority class of the original learning data” which is the upper limit value of the range of the weight N. (N ′ = N _r ), the number of learning data belonging to the minority class of new learning data is equal to the number of learning data belonging to the majority class of new learning data. Further, when the value of the temporary weight N ′ is equal to “1” which is the lower limit value of the range of the weight N (N ′ = 1), N ′ ₊ = N ′ ₋ = 1, and therefore oversampling is also under. Sampling will not be performed. Therefore, regardless of the value of the priority r, the range of the weight N is the same as when only oversampling described in the first embodiment is performed or only undersampling described in the second embodiment is performed. What is necessary is just to set by "1"-" _Nr ".

Furthermore, when creating a classifier using the optimum value N _opt weights, (25), for a few classes by substituting the optimum value N _opt weights instead of "N'" in (26) Weight N ₊ and a majority class weight N _−, and after oversampling the minority class learning data using the minority class weight N ₊ , the majority class weight N ₋ The class learning data may be undersampled.
As described above, the effects described in the first and second embodiments can be obtained. Also in this embodiment, various modifications described in the first and second embodiments can be adopted.

In the present embodiment, the case where the priority r for increasing the learning data belonging to the minority class is used with respect to the decrease in the number of learning data belonging to the majority class has been described as an example. However, the priority r of decreasing the number of learning data belonging to the majority class with respect to the increase of learning data belonging to the minority class may be used. In such a case, the equations (29) and (30) may be used instead of the equations (25) and (26).
_{N ′ +} = _{N ′} / _{N ′ −} (29)
N ′ ₋ = r × N ′ + (1−r) (30)

The embodiment of the present invention described above can be realized by a computer executing a program. Further, a computer-readable recording medium in which the program is recorded and a computer program product such as the program can also be applied as an embodiment of the present invention. As the recording medium, for example, a flexible disk, a hard disk, an optical disk, a magneto-optical disk, a CD-ROM, a magnetic tape, a nonvolatile memory card, a ROM, or the like can be used.
In addition, the embodiments of the present invention described above are merely examples of implementation in carrying out the present invention, and the technical scope of the present invention should not be construed as being limited thereto. Is. That is, the present invention can be implemented in various forms without departing from the technical idea or the main features thereof.

[Relationship with Claims]
<Claims 1 and 9>
The temporary weight determination unit is realized by using, for example, the weight range input unit 102 and the temporary weight determination unit 103, and the temporary weight determination step is realized by performing, for example, steps S502 and S503. Here, the lower limit value corresponds to N _min , for example, and the upper limit value corresponds to N _max .
The learning data sampling means is realized, for example, by using the learning data sampling unit 104, and the learning data sampling step is realized, for example, by performing the process of step S505 (FIG. 6).
The learning means is realized, for example, by using the learning unit 105, and the learning process is realized, for example, by performing the process of step S506.
The evaluation value calculation means is realized by using, for example, the evaluation value calculation unit 106, and the evaluation value calculation process is realized by, for example, performing the process of step S507.
The optimum weight derivation means is realized by using, for example, the optimum weight derivation unit 107, and the optimum weight derivation step is realized by, for example, performing the process of step S509 (FIG. 7).
The classifier creating means is realized by using, for example, the learning data sampling unit 108 and the learning unit 109, and the classifier creating process is realized by performing, for example, steps S409 to S411.
<Claims 3 and 11>
Claims 3 and 11 correspond to, for example, the learning data sampling process in the first embodiment.
<Claims 4 and 12>
Claims 4 and 12 correspond to, for example, the learning data sampling process in the second embodiment.
<Claims 5 and 13>
Claims 5 and 13 correspond to, for example, the learning data sampling processing in the third embodiment.

DESCRIPTION OF SYMBOLS 100 Classifier preparation apparatus 101 Learning data input part 102 Weight range input part 103 Temporary weight determination part 104 Learning data sampling part 105 Learning part 106 Evaluation value calculation part 107 Optimal weight calculation part 108 Learning data sampling part 109 Learning part 110 Classifier storage Part

Claims (17)

Among the learning data that is known to belong to which of the two classes, increasing the number of learning data of a minority class, which is a class to which a relatively small number of learning data belongs, by a factor corresponding to the weight, The number of learning data is changed by at least one of reducing the number of learning data of a large number of classes, which is a class to which a large number of learning data belongs, by a factor corresponding to the inverse of the weight. Creating a new learning data, and using the new learning data, a classifier creating device for creating a classifier for determining which of two classes the given data belongs to,
The upper limit value and the lower limit value of the weight are determined from a range of one or more values less than or equal to a value obtained by dividing the number of learning data of the majority class by the number of learning data of the minority class, and the upper limit value and the lower limit value. Temporary weight determining means for determining a plurality of temporary weights having different values from a range of values;
At least one of increasing the number of learning data of the minority class by a factor corresponding to the provisional weight and reducing the number of learning data of the majority class by a factor corresponding to the inverse of the provisional weight. Learning data sampling means for performing one for each of the plurality of temporary weights to create a plurality of sets of the new learning data;
Using the new learning data after the number of learning data has been changed by the learning data sampling means, a temporary classifier for determining which of the two classes the given learning data belongs is created Learning means for performing each of the new learning data sets to obtain a plurality of the temporary classifiers;
Calculating an evaluation value for evaluating the performance of the temporary classifier obtained by the learning means based on the result of classifying the learning data into one of the two classes by the temporary classifier; An evaluation value calculation means for obtaining a plurality of evaluation values by performing for each of a plurality of temporary classifiers;
Using the evaluation value obtained by the evaluation value calculation means and the provisional weight used when obtaining the evaluation value, the relationship between the evaluation value and the weight is obtained, and in the obtained relationship, from the lower limit value Optimum weight deriving means for deriving a weight corresponding to the evaluation value having the largest value in the range up to the upper limit value as an optimum value of the weight;
Using the optimum value of the weight as the weight, increasing the number of learning data of the minority class by a factor corresponding to the weight, and increasing the number of learning data of the majority class by a factor corresponding to the inverse of the weight A classification for determining whether the given learning data belongs to which of the two classes using the new learning data by creating at least one of reducing and creating the new learning data A classifier creating means for creating a container;
A classifier creating apparatus comprising:   The classifier creating apparatus according to claim 1, wherein the number of temporary weights is 3 or more. The learning data sampling means includes at least one of the learning data of the minority class such that the number of new learning data of the minority class is equal to the number of learning data of the minority class multiplied by the temporary weight. Copying a part for each of the plurality of temporary weights to create a plurality of sets of the new learning data,
The classifier creating unit replicates the learning data of the minority class so that the number of new learning data of the minority class is equal to the number of learning data of the minority class multiplied by the optimum value of the weight. The classifier creating apparatus according to claim 1, wherein new learning data is generated, and the classifier is generated using the generated new learning data. The learning data sampling means is configured so that the number of new learning data of the majority class is equal to the number of learning data of the majority class multiplied by the inverse of the temporary weight. Deleting a part for each of the plurality of temporary weights to create a plurality of sets of the new learning data;
The classifier creating means includes the learning data of the multi-class so that the number of the new learning data of the multi-class is the number of the learning data of the multi-class multiplied by the inverse of the optimum value of the weight. The classifier creating apparatus according to claim 1, wherein a part of the data is deleted to generate new learning data, and the classifier is generated using the generated new learning data. Priority of increasing the number of learning data of the minority class with respect to a decrease in the number of learning data of the majority class, or priority of decreasing the number of learning data of the majority class with respect to an increase in the number of learning data of the minority class A priority that takes a value between 0 and 1 in advance, and using the temporary weight, a temporary weight for a minority class, which is a temporary weight for the learning data of the minority class, and the majority Provisional weight deriving means for deriving provisional weights for a large number of classes, which are provisional weights for the learning data of the class;
Using the priority and the optimum value of the weight, a weight for the minority class that is a weight for the learning data of the minority class and a weight for the minority class that is a weight for the learning data of the majority class are derived. Weight derivation means for
The learning data sampling unit is configured to learn the minority class so that the number of new learning data of the minority class is equal to the number of learning data of the minority class multiplied by a provisional weight for the minority class. Duplicating at least a part of the data, and the number of new learning data of the majority class is equal to the number of learning data of the majority class multiplied by the inverse of the provisional weight for the majority class. Deleting a part of the learning data of the large number of classes for each of the plurality of temporary weights to create a plurality of sets of the new learning data,
The classifier creating means is configured so that the number of new learning data of the minority class is equal to the number of learning data of the minority class multiplied by the optimum value of the weight for the minority class. Duplicating learning data, and the number of new learning data of the majority class is equal to the number of learning data of the majority class multiplied by the inverse of the optimum value of the weight for the majority class. Deleting a part of the learning data of a large number of classes, generating new learning data, creating the classifier using the generated new learning data, the priority being the learning data of the large number of classes If the priority value is 1, the provisional weight of the minority class becomes the provisional weight when the priority value is 1. Be the same And when the priority value is 0, the provisional weights of the multiple classes are the same as the provisional weights,
When the priority is a priority of decreasing the number of learning data of the majority class with respect to an increase of the number of learning data of the minority class, when the priority value is 1, the majority class The provisional weight of the minority class is the same as the provisional weight when the provisional weight is the same as the provisional weight and the priority value is 0. The classifier creating apparatus according to claim 1 or 2.   The evaluation value calculating means includes learning data that actually belong to a minority class among learning data classified into a minority class based on a result of classifying the learning data into one of the two classes by the temporary classifier. Calculate the relevance rate, which represents the proportion of data included, and the recall, which is the proportion of learning data that actually belongs to the minority class, among the learning data that belongs to the minority class. The classifier creating apparatus according to claim 1, wherein an F value that is a weighted harmonic average of a rate and a recall rate is calculated as the evaluation value. The temporary weight determining means derives three temporary weights each having the upper limit value, the lower limit value, and a value that is half of the sum of the upper limit value and the lower limit value,
7. The optimum weight deriving unit obtains a relationship between the evaluation value and the weight on the assumption that the evaluation value is expressed by a quadratic function of the weight. The classifier preparation apparatus described in 1.   The classifier creating apparatus according to claim 1, wherein the classifier is a decision tree. Among the learning data that is known to belong to which of the two classes, increasing the number of learning data of a minority class, which is a class to which a relatively small number of learning data belongs, by a factor corresponding to the weight, The number of learning data is changed by at least one of reducing the number of learning data of a large number of classes, which is a class to which a large number of learning data belongs, by a factor corresponding to the inverse of the weight. Creating a new learning data, and using the new learning data, a classifier creating method for creating a classifier for determining which of two classes the given data belongs to,
The upper limit value and the lower limit value of the weight are determined from a range of one or more values less than or equal to a value obtained by dividing the number of learning data of the majority class by the number of learning data of the minority class, and the upper limit value and the lower limit value. A provisional weight determination step for determining a plurality of provisional weights having different values from a range of values;
At least one of increasing the number of learning data of the minority class by a factor corresponding to the provisional weight and reducing the number of learning data of the majority class by a factor corresponding to the inverse of the provisional weight. A learning data sampling step of performing one for each of the plurality of temporary weights to create a plurality of sets of the new learning data;
Using the new learning data after the number of learning data has been changed by the learning data sampling step, a temporary classifier for determining which of the two classes the given learning data belongs is created Performing a learning step for each set of new learning data to obtain a plurality of temporary classifiers;
Calculating an evaluation value for evaluating the performance of the temporary classifier obtained by the learning step based on the result of classifying the learning data into one of the two classes by the temporary classifier; An evaluation value calculation step for obtaining a plurality of evaluation values by performing for each of a plurality of temporary classifiers;
Using the evaluation value obtained in the evaluation value calculation step and the temporary weight used when obtaining the evaluation value, the relationship between the evaluation value and the weight is obtained, and in the obtained relationship, from the lower limit value An optimum weight derivation step for deriving a weight corresponding to the evaluation value having the largest value in the range up to the upper limit value as an optimum value of the weight;
Using the optimum value of the weight as the weight, increasing the number of learning data of the minority class by a factor corresponding to the weight, and increasing the number of learning data of the majority class by a factor corresponding to the inverse of the weight A classification for determining whether the given learning data belongs to which of the two classes using the new learning data by creating at least one of reducing and creating the new learning data A classifier creation process for creating a container;
A classifier creating method characterized by comprising:   The classifier creation method according to claim 9, wherein the number of provisional weights is 3 or more. The learning data sampling step includes at least one of the learning data of the minority class such that the number of new learning data of the minority class is equal to the number of learning data of the minority class multiplied by the temporary weight. Copying a part for each of the plurality of temporary weights to create a plurality of sets of the new learning data,
The classifier creating step replicates the learning data of the minority class so that the number of new learning data of the minority class is equal to the number of learning data of the minority class multiplied by the optimum value of the weight. The classifier creation method according to claim 9 or 10, wherein new learning data is generated and the classifier is generated using the generated new learning data. In the learning data sampling step, the number of the learning data of the majority class is adjusted so that the number of the new learning data of the majority class is the number of the learning data of the majority class multiplied by the reciprocal of the temporary weight. Deleting a part for each of the plurality of temporary weights to create a plurality of sets of the new learning data;
In the classifier creating step, the learning data of the multi-class is set so that the number of new learning data of the multi-class is the number obtained by multiplying the number of learning data of the multi-class by the inverse of the optimum value of the weight. The classifier creation method according to claim 9 or 10, wherein a part of is deleted to generate new learning data, and the classifier is generated using the generated new learning data. Priority of increasing the number of learning data of the minority class with respect to a decrease in the number of learning data of the majority class, or priority of decreasing the number of learning data of the majority class with respect to an increase in the number of learning data of the minority class A priority that takes a value between 0 and 1 in advance, and using the temporary weight, a temporary weight for a minority class, which is a temporary weight for the learning data of the minority class, and the majority A provisional weight derivation step for deriving provisional weights for multiple classes, which are provisional weights for the learning data of the class;
Using the priority and the optimum value of the weight, a weight for the minority class that is a weight for the learning data of the minority class and a weight for the minority class that is a weight for the learning data of the majority class are derived. A weight deriving step to
In the learning data sampling step, the learning of the minority class is performed so that the number of new learning data of the minority class is equal to the number of the learning data of the minority class multiplied by the provisional weight for the minority class. Duplicating at least a part of the data, and the number of new learning data of the majority class is equal to the number of learning data of the majority class multiplied by the inverse of the provisional weight for the majority class. Deleting a part of the learning data of the large number of classes for each of the plurality of temporary weights to create a plurality of sets of the new learning data,
In the classifier creating step, the number of new learning data in the minority class is equal to the number obtained by multiplying the number of learning data in the minority class by the optimum value of the weight for the minority class. Duplicating learning data, and the number of new learning data of the majority class is equal to the number of learning data of the majority class multiplied by the inverse of the optimum value of the weight for the majority class. Deleting a part of the learning data of many classes, generating new learning data, creating the classifier using the generated new learning data,
When the priority is a priority of an increase in the number of learning data of the minority class with respect to a decrease in the number of learning data of the majority class, when the priority value is 1, the minority class The temporary weights of the multiple classes are the same as the temporary weights, and when the priority value is 0,
When the priority is a priority of decreasing the number of learning data of the majority class with respect to an increase of the number of learning data of the minority class, when the priority value is 1, the majority class The provisional weight of the minority class is the same as the provisional weight when the provisional weight is the same as the provisional weight and the priority value is 0. The classifier creation method according to claim 9 or 10.   In the evaluation value calculating step, learning actually belonging to a minority class among learning data classified into a minority class based on a result of classifying the learning data into one of the two classes by the temporary classifier Calculate the relevance rate, which represents the proportion of data included, and the recall, which is the proportion of learning data that actually belongs to the minority class, among the learning data that belongs to the minority class. The classifier creation method according to claim 9, wherein an F value that is a weighted harmonic average of a rate and the recall rate is calculated as the evaluation value. The provisional weight determination step derives three provisional weights each having the upper limit value, the lower limit value, and a value that is ½ of the sum of the upper limit value and the lower limit value,
15. The optimum weight derivation step obtains a relationship between the evaluation value and the weight on the assumption that the evaluation value is expressed by a quadratic function of the weight. The classifier creation method described in 2.   The classifier creation method according to claim 9, wherein the classifier is a decision tree.   A computer program for causing a computer to function as each means of the classifier creating apparatus according to claim 1.

Images