Outlier removal method and device

Abstract

To provide an outlier removal method and device capable of appropriately removing an outlier.SOLUTION: An outlier removal method for removing an outlier included in learning data 31 includes: a prediction error calculation step of dividing the learning data 31 into teacher data and test data, creating a regression model by using the teacher data, and repeating, a predetermined number of times, calculation of a prediction error relative to the created regression model by using the test data; a distribution calculation step of extracting, for each data included in the learning data 31, a prediction error when the data is used as test data from the prediction error obtained in the prediction error calculation step, and obtaining, for each data, an index value that characterizes a distribution of the extracted prediction error; an outlier determination step of determining whether or not each data is an outlier based on an index value of each data obtained in the distribution calculation step; and an outlier removal step of removing data determined as the outlier in the outlier determination step.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a method and apparatus for removing outliers. [Background technology]

[0002] There are known methods for making various predictions using machine learning. For example, when predicting the physical properties of a material with an unknown composition, machine learning is performed using data already obtained by test manufacturing or the like as learning data (teacher data), the correlation between the composition of the material and its physical properties is learned, and predictions are made using a regression model obtained as a result of the learning.

[0003] Prior art document information related to the invention of this application includes Patent Document 1. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2020-123365 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, if the learning data contains outliers, which are erroneous data or data with large errors, the prediction accuracy of the regression model obtained using the learning data will decrease. Therefore, it is desirable to remove the outliers from the learning data prior to machine learning. However, it has been difficult to properly determine which data is an outlier. In particular, when the learning data is sparse data, it has been difficult to properly remove the outliers.

[0006] Therefore, an object of the present invention is to provide an outlier removal method and device capable of appropriately removing outliers. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems, the present invention provides a method for removing outliers contained in learning data including data on explanatory variables and dependent variables used in machine learning, the method comprising: a prediction error calculation step of dividing the learning data into teacher data and test data, creating a regression model showing a correlation between the explanatory variables and the dependent variables using the teacher data, and repeating a predetermined number of times a calculation of a prediction error for the created regression model using the test data; a distribution calculation step of extracting, from the prediction errors obtained in the prediction error calculation step, a prediction error for each data item included in the learning data when the data item is used as the test data, and determining, for each data item included in the learning data, an index value characterizing a distribution of the extracted prediction errors; an outlier determination step of determining whether each data item is an outlier based on the index value for each data item determined in the distribution calculation step; and an outlier removal step of removing data determined to be an outlier in the outlier determination step.

[0008] Moreover, for the purpose of solving the above-mentioned problems, the present invention provides an apparatus for removing outliers included in learning data including data on explanatory variables and objective variables used in machine learning, the outlier removal device comprising: a prediction error calculation processing unit that divides the learning data into teacher data and test data, creates a regression model showing a correlation between the explanatory variables and the objective variable using the teacher data, and repeats a predetermined number of times calculating a prediction error for the created regression model using the test data; a distribution calculation processing unit that extracts, from the prediction errors obtained by the prediction error calculation processing unit, a prediction error when each data item included in the learning data is used as the test data, and obtains, for each data item included in the learning data, an index value characterizing a distribution of the extracted prediction errors; an outlier determination processing unit that determines whether each data item is an outlier based on the index value for each data item obtained by the distribution calculation processing unit; and an outlier removal processing unit that removes data determined to be an outlier by the outlier determination processing unit. Effect of the Invention

[0009] According to the present invention, it is possible to provide an outlier removal method and device capable of appropriately removing outliers. [Brief description of the drawings]

[0010] [Figure 1] 1 is a schematic configuration diagram of an outlier removal device according to an embodiment of the present invention. [Diagram 2] FIG. 11 is a diagram illustrating an example of learning data. [Diagram 3] FIG. 1A is a diagram for explaining a prediction error calculation process, and FIG. 1B is a diagram for explaining a distribution calculation process. [Figure 4] FIG. 13 is a diagram for explaining the reason why both MAE and MAPE are used as index values. [Diagram 5] FIG. 13 is a diagram showing an example of distribution of the medians of MAE and MAPE obtained by distribution calculation processing. [Figure 6] FIG. 2 is a flow diagram of an outlier removal method according to an embodiment of the present invention. [Figure 7] FIG. 11 is a flow diagram of a prediction error calculation process. [Figure 8] FIG. 13 is a flow diagram of a distribution calculation process. [Figure 9] FIG. 11 is a flow diagram of an outlier determination process. [Figure 10] FIG. 11 is a flow diagram of an outlier removal process. [Figure 11] (a) is a graph showing the change in the coefficient of determination when outliers are removed, (b) is a graph showing the change in the MAE when outliers are removed, and (c) is a graph showing the change in the MAPE when outliers are removed. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] [Embodiment Mode] Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings.

[0012] 1 is a schematic diagram of an outlier removal device 1 according to the present embodiment. The outlier removal device 1 is a device that removes outliers contained in training data 31 used for machine learning. An outlier means a value that is significantly different from other data due to, for example, a measurement error, a human error such as an incorrect reading of an instrument or an input error, or the influence of noise. By detecting and removing outliers from the training data 31, it is expected that the prediction accuracy will be improved when machine learning is performed using the training data 31.

[0013] The outlier removal device 1 has a control unit 2 and a storage unit 3. The outlier removal device 1 is, for example, a computer such as a personal computer or a server device, and includes a calculation element such as a CPU, memories such as a RAM and a ROM, a storage device such as a hard disk, and a communication interface which is a communication device such as a LAN card.

[0014] The control unit 2 has a data acquisition processing unit 21, a prediction error calculation processing unit 22, a distribution calculation processing unit 23, an outlier determination processing unit 24, and an outlier removal processing unit 25. Details of each unit will be described later. The storage unit 3 is realized by a predetermined storage area of ​​a memory or a storage device.

[0015] The outlier removal device 1 also has a display 4 and an input device 5. The display 4 is, for example, a liquid crystal display, and the input device 5 is, for example, a keyboard or a mouse. The display 4 may be configured as a touch panel and also function as the input device 5. The display 4 and the input device 5 may also be configured separately from the outlier removal device 1 and configured to be able to communicate with the outlier removal device 1 via wireless communication or the like. In this case, the display 4 or the input device 5 may be configured as a mobile terminal such as a tablet or a smartphone.

[0016] (Data acquisition processing unit 21) The data acquisition processing unit 21 performs a data acquisition process to acquire learning data 31 from an external device. In the data acquisition process, the learning data 31 is acquired via a network, for example, from a prediction device or the like that uses the learning data 31. Note that the learning data 31 may be input to the outlier removal device 1 via a medium such as a USB memory, and the method of acquiring the learning data 31 is not particularly limited.

[0017] (Learning data 31) Here, the learning data 31 will be explained. FIG. 2 is a diagram showing an example of the learning data 31. The learning data 31 is a database used as teacher data during machine learning, and includes data on explanatory variables and objective variables used in machine learning. In the example of FIG. 2, the blend amounts of materials such as polymers and fillers are used as explanatory variables, and the physical properties (here, tear strength) of a composite material manufactured using the materials are used as the objective variable. By performing machine learning using this learning data 31 and creating a regression model showing the correlation between the explanatory variables (blended amounts of each material) and the objective variable (physical properties), it becomes possible to predict the physical properties obtained when a composite material is manufactured with an unknown blend of materials.

[0018] (Prediction error calculation processing unit 22) The prediction error calculation processing unit 22 performs a prediction error calculation process in which the learning data 31 is divided into teacher data and test data, a regression model (trained model) showing the correlation between explanatory variables and objective variables is created using the teacher data, and a prediction error is calculated for the created regression model using the test data, which is repeated a predetermined number of times. The prediction error calculation process corresponds to the prediction error calculation step of the present invention.

[0019] As shown in FIG. 3(a), the prediction error calculation processing unit 22 first randomly divides the learning data 31 into teacher data and test data. The ratio of teacher data to test data may be set in advance, for example, 70% teacher data and 30% test data. After the data division, machine learning is performed using the teacher data, and a regression model showing the correlation between the explanatory variables and the objective variable is created as a result of the machine learning. Then, the test data is applied to the created regression model to obtain a predicted value of the objective variable and a prediction error showing the error between the predicted value and the actual value (the value of the objective variable in the test data).

[0020] As the prediction error, the mean error (ME), the mean absolute error (MAE), the mean squared error (RMSE), the mean percent error (MPE), the mean absolute error (MAPE), the root mean squared error (RMSPE), etc. can be used. More preferably, as the prediction error, at least one of the ME, MAE, and RMSE and one of the MPE, MAPE, and RMSPE are used. In this embodiment, the MAE and MAPE are used as the prediction errors.

[0021] FIG. 4 shows an example of the relationship between the predicted value and the actual value when the test data is applied to the regression model created using the teacher data. As shown in FIG. 4, when only the MAE (or ME, RMSE) is used, there is a sufficient margin when the value of the objective variable (here, tear strength) is small, but when the value of the objective variable (here, tear strength) is large, the judgment becomes strict, and there is a risk that data that is not an outlier is judged as an outlier. On the other hand, when only the MAPE (or MPE, RMSPE) is used, there is a sufficient margin when the value of the objective variable (here, tear strength) is large, but when the value of the objective variable (here, tear strength) is small, the judgment becomes strict, and there is a risk that data that is not an outlier is judged as an outlier. Therefore, in this embodiment, both the MAE (or ME, RMSE) and the MAPE (or MPE, RMSPE) are used as the prediction error. The obtained prediction error is stored in the storage unit 3 as the prediction error data 32.

[0022] The prediction error calculation processing unit 22 repeats the division of the learning data 31, the creation of a regression model using the teacher data, and the calculation of the prediction error by applying the test data to the regression model a preset number of times. The number of repetitions may be appropriately determined according to the number of data included in the learning data 31, and may be appropriately determined so that several tens to several hundreds (preferably 100 or more) of prediction errors (here, MAE and MAPE) using any data as test data are obtained. Here, the number of repetitions is set to 955, taking into consideration that the number of data is 212. As a result, a prediction error of about 300 can be obtained for each data when the data is used as test data.

[0023] (Distribution calculation processing unit 23) The distribution calculation processing unit 23 performs distribution calculation processing to extract, for each data included in the learning data 31, a prediction error when that data is used as test data from the prediction errors (MAE and MAPE) obtained in the prediction error calculation processing, and to obtain an index value characterizing the distribution of the extracted prediction errors for each data included in the learning data 31. The distribution calculation processing corresponds to a distribution calculation step of the present invention.

[0024] In the distribution calculation process, as shown in FIG. 3(b), for example, MAE and MAPE when data No. 1 is used as test data are extracted from the prediction error data 32, and the distribution of MAE and the distribution of MAPE are obtained, and an index value characterizing the distribution is calculated. In this embodiment, the median of the distribution is used as the index value. This is because, by using the median as the index value, the number of data exceeding the index value in the error distribution for data determined to be an outlier is always half (50%) or more, and it is easy to determine how much the data exceeds the index value. However, this is not limiting, and for example, the average value or the like may be used.

[0025] In addition, in the distribution calculation process, the proportion of the distribution of the prediction errors that is equal to or greater than a preset judgment criterion value is calculated. More specifically, the proportion of the MAE (or ME, RMSE) that is equal to or greater than a preset first judgment criterion value (3N / mm in FIG. 3(b)) and the proportion of the MAPE (or MPE, RMSPE) that is equal to or greater than a preset second judgment criterion value (30% in FIG. 3(b)) are calculated, and the average value of these proportions (hereinafter referred to as the criterion exceedance rate) is calculated. For example, if 90 out of 100 data items in MAE are equal to or greater than the first judgment criterion value and 100 out of 100 data items in MAPE are equal to or greater than the second judgment criterion value, the criterion exceedance rate is (90 / 100+100 / 100) / 2=0.95.

[0026] In the distribution calculation process, the median of the MAE distribution, the median of the MAPE distribution, and the benchmark exceedance rate are calculated for all data (212 data in this case) included in the learning data 31, and these calculation results are stored in the memory unit 3 as distribution data 33.

[0027] (Outlier determination processing unit 24) The outlier determination processing unit 24 performs an outlier determination process to determine whether each data is an outlier based on the index value of each data (here, the median of the distribution) calculated by the distribution calculation process. The outlier determination process corresponds to the outlier determination step of the present invention.

[0028] The outlier determination processor 24 determines that data whose index value (here, the median of the distribution) is equal to or greater than a preset judgment reference value is an outlier. More specifically, the outlier determination processor 24 determines that data whose median of MAE (or ME, RMSE) is equal to or greater than a preset first judgment reference value (e.g., 3 N / mm) and whose median of MAPE (or MPE, RMSPE) is equal to or greater than a preset second judgment reference value (e.g., 30%) is an outlier. The number (No.) of the data determined to be an outlier is stored in the storage unit 3 as outlier candidate data 34.

[0029] Fig. 5 is a diagram showing an example of the distribution of the median values ​​of MAE and MAPE obtained by distribution calculation processing. When the first judgment criterion value is set to 3 N / mm and the second judgment criterion value is set to 30%, the data plotted in the hatched area in Fig. 5 is judged to be an outlier.

[0030] (Outlier removal processing unit 25) The outlier removal processing unit 25 performs an outlier removal process for removing data determined to be an outlier in the outlier determination process. The outlier removal process corresponds to an outlier removal step of the present invention.

[0031] The outlier removal processor 25 is configured to remove only one of a plurality of data items determined to be outliers in the outlier determination process. This is because the data items determined to be outliers may include data items with large prediction errors due to the influence of other outliers, and this is to prevent data items that are not outliers from being removed.

[0032] When there are multiple pieces of data determined to be outliers in the outlier determination process, the outlier removal processor 25 removes only the data having the largest proportion of data exceeding the judgment reference value in the distribution of prediction errors. More specifically, the outlier removal processor 25 removes only one piece of data having the largest criteria exceedance rate from the learning data 31 among the pieces of data included in the outlier candidate data 34 (removes the outliers and updates the learning data 31). As described above, the criteria exceedance rate is the average value of the proportion of MAE (or ME, RMSE) that is equal to or greater than the first judgment reference value and the proportion of MAPE (or MPE, RMSPE) that is equal to or greater than the second judgment reference value. The outlier removal processor 25 stores the data removed as an outlier in the storage unit 3 as the outlier data 35. The criteria exceedance rate may be set as the product of the proportion of data that is equal to or greater than the first judgment reference value and the proportion of data that is equal to or greater than the second judgment reference value.

[0033] The control unit 2 repeats the prediction error calculation process, the outlier determination process, and the outlier removal process until there is no data determined to be an outlier in the outlier determination process. Then, when there is no data determined to be an outlier in the outlier determination process, the control unit 2 ends the process. In this way, it is possible to obtain the learning data 31 from which the outliers have been removed.

[0034] (Outlier removal method) Fig. 6 is a flow diagram of the outlier removal method according to the present embodiment. As shown in Fig. 6, first, in step S1, the data acquisition processing unit 21 performs a data acquisition process to acquire learning data 31 from an external device or the like. The acquired learning data 31 is stored in the storage unit 3.

[0035] Then, in step S2, a prediction error calculation process is performed. In the prediction error calculation process, as shown in FIG. 7, first, in step S1, an initial value of 1 is substituted into a variable n representing the number of repetitions. Also, a variable n representing the maximum number of repetitions is set to 0. max Then, in step S22, the prediction error calculation processing unit 22 randomly divides the learning data 31 into teacher data and test data. Then, in step S23, the prediction error calculation processing unit 22 creates a regression model using the teacher data. Then, in step S24, the prediction error calculation processing unit 22 applies the test data to the created regression model to find the prediction errors, MAE and MAPE. Then, in step S25, the prediction error calculation processing unit 22 stores the found MAE and MAPE in the memory unit 3 as prediction error data 32. Then, in step S26, the variable n is set to n max If the determination in step S26 is NO (N), n is incremented in step S27 and the process returns to step S22. If the determination in step S26 is YES (Y), the process returns and proceeds to step S3 in FIG.

[0036] In step S3, a distribution calculation process is performed. In the distribution calculation process, as shown in FIG. 8, first, in step S31, the distribution calculation processing unit 23 extracts the MAE and MAPE from the prediction error data 32 for each data included in the learning data 31 when the data is used as test data. Thereafter, in step S32, the distribution calculation processing unit 23 calculates the median of the distribution of the MAE and the median of the distribution of the MAPE for each data. Then, in step S33, the distribution calculation processing unit 23 calculates the average value of the ratio of the MAE being equal to or greater than a first judgment reference value set in advance and the ratio of the MAPE being equal to or greater than a second judgment reference value set in advance, that is, the standard exceedance rate, for each data. Thereafter, in step S34, the distribution calculation processing unit 23 stores the median and standard exceedance rate of the MAE and MAPE of each data calculated by the calculation in the storage unit 3 as distribution data 33. Thereafter, the process returns and proceeds to step S4 in FIG. 6.

[0037] In step S4, an outlier determination process is performed. In the outlier determination process, as shown in FIG. 9, first, in step S41, the outlier determination processing unit 24 assigns an initial value of 1 to a variable i indicating the number of data. Thereafter, in step S42, the outlier determination processing unit 24 determines whether the MAE of the i-th data is equal to or greater than a first determination criterion value. If the determination is NO in step S42, in step S43, the outlier determination processing unit 24 determines that the i-th data is not an outlier, and proceeds to step S46. If the determination is YES in step S42, in step S44, the outlier determination processing unit 24 determines whether the MAPE of the i-th data is equal to or greater than a second determination criterion value. If the determination is NO in step S44, in step S43, the outlier determination processing unit 24 determines that the i-th data is not an outlier, and proceeds to step S46. If the determination is YES in step S44, in step S45, the outlier determination processing unit 24 determines that the i-th data is an outlier, and proceeds to step S46.

[0038] In step S46, the outlier determination processor 24 stores the determination results from steps S43 and S45 in the storage unit 3. Here, the outlier determination processor 24 stores the number i of the data determined to be an outlier in the storage unit 3 as outlier candidate data 34. Thereafter, in step S47, the outlier determination processor 24 determines whether determination has been performed on all data. If the determination in step S47 is NO, the variable i is incremented in step S48, and the process returns to step S42. If the determination in step S47 is YES, the process returns and proceeds to step S5 in FIG. 6.

[0039] In step S5, it is determined whether or not there is an outlier as a result of the determination in step S4. If the determination in step S5 is NO, the process is terminated. If the determination in step S6 is YES, the outlier removal process is performed in step S6. In the outlier removal process, as shown in FIG. 10, in step S61, the outlier removal processor 25 refers to the distribution data 33 and extracts data having the largest standard exceedance rate from among the data included in the outlier candidate data 34. Then, in step S62, the outlier removal processor 25 deletes the extracted data to update the learning data 31, and stores the data removed from the learning data 31 in the storage unit 3 as outlier data 35. Then, the process returns to step S2 in FIG. 6.

[0040] (Changes in prediction accuracy due to removal of outliers) We investigated the change in prediction accuracy when outliers were removed using the outlier removal method according to the present embodiment. As the learning data 31, data containing 212 pieces of data was used, and as explanatory variables, the compounding amounts of 33 kinds of materials and material information in which the characteristics of the material composition (e.g., the volume ratio of filler) were quantified were used, and as the objective variable, tear strength was used.

[0041] The data division of the learning data 31, the creation of a regression model using the teacher data obtained by the division, and the measurement of the prediction error (MAE and MAPE) by applying the test data obtained by the division to the regression model were repeated 955 times, and the prediction error of about 300 was obtained for each data when the data was used as test data. Then, the median of the distribution of MAE, the median of the distribution of MAPE, and the standard exceedance rate were calculated from the distribution of the prediction error of each data, and when the median of the distribution of MAE was 3N / mm or more and the median of the distribution of MAPE was 30% or more, it was determined to be an outlier. Of the data determined to be an outlier, the data with the largest standard exceedance rate was removed, and the process was repeated until there was no data determined to be an outlier. As a result, there was no data determined to be an outlier after removing 10 outliers.

[0042] Each time an outlier is removed, the coefficient of determination R 2 , MAE, and MAPE were evaluated. The results are shown in Figure 11(a) to (c). As shown in Figure 11(a), the coefficient of determination R 2 It can be seen that the more outliers are removed, the higher the MAE and MAPE increase. In addition, as shown in Figures 11(b) and 11(c), the more outliers are removed, the lower the MAE and MAPE decrease, confirming that the prediction accuracy is improved.

[0043] (Functions and Effects of the Embodiments) As described above, the outlier removal method according to this embodiment includes a prediction error calculation step of dividing the training data 31 into teacher data and test data, creating a regression model showing the correlation between explanatory variables and target variables using the teacher data, and repeating the process a predetermined number of times of calculating prediction errors for the created regression model using the test data; a distribution calculation step of extracting, from the prediction errors obtained in the prediction error calculation step, prediction errors for each data item included in the training data 31 when that data item is used as test data, and determining, for each data item included in the training data 31, an index value characterizing the distribution of the extracted prediction errors (here, the median); an outlier determination step of determining whether each data item is an outlier based on the index value for each data item determined in the distribution calculation step; and an outlier removal step of removing data determined to be an outlier in the outlier determination step.

[0044] This makes it possible to appropriately remove outliers when they are included in the learning data 31. As a result, it becomes possible to use the learning data 31 from which the outliers have been appropriately removed for machine learning, thereby making it possible to improve the prediction accuracy of physical property predictions and the like.

[0045] In this embodiment, data that has a large prediction error (i.e., data that cannot be predicted well by the created regression model) when applied to a regression model created from teacher data divided from the learning data 31 is removed as an outlier, but if there is only one regression model for determining outliers, there is a risk that a non-outlier may be determined to be an outlier by chance. By using a method of dividing the learning data 31 multiple times to create multiple different regression models and evaluating the distribution of prediction errors when each data is applied as test data as in this embodiment, the possibility of a non-outlier being determined to be an outlier by chance as described above can be significantly reduced, and it becomes possible to properly remove outliers even if the learning data 31 is sparse data.

[0046] (Modification) Although not mentioned in the above embodiment, the outlier removal device 1 may be incorporated as a function in a prediction device that predicts physical properties and the like using the learning data 31. In this case, the prediction device is provided with a regression model creation unit that performs machine learning using the learning data 31 from which the outliers have been removed and creates a regression model indicating the correlation between the explanatory variables and the objective variable, and a prediction unit that predicts physical properties and the like using the regression model created by the regression model creation unit.

[0047] (Summary of the embodiment) Next, the technical ideas grasped from the above-described embodiment will be described by using the reference numerals and the like in the embodiment. However, the reference numerals and the like in the following description do not limit the components in the claims to the members and the like specifically shown in the embodiment.

[0048] [1] A method for removing outliers contained in training data (31) including data on explanatory variables and objective variables used in machine learning, the method comprising: a prediction error calculation step of dividing the training data (31) into teacher data and test data, creating a regression model showing a correlation between the explanatory variables and the objective variable using the teacher data, and repeating a predetermined number of times a calculation of a prediction error for the created regression model using the test data; a distribution calculation step of extracting, from the prediction errors obtained in the prediction error calculation step, a prediction error for each data item included in the training data (31) when the data item is used as the test data, and determining, for each data item included in the training data (31), an index value characterizing a distribution of the extracted prediction errors; an outlier determination step of determining whether each data item is an outlier based on the index value for each data item determined in the distribution calculation step; and an outlier removal step of removing data determined to be an outlier in the outlier determination step.

[0049] [2] The outlier removal method according to [1], wherein in the outlier removal step, when there are multiple data determined to be outliers in the outlier determination step, only one of the data is removed, and the prediction error calculation step, the outlier determination step, and the outlier removal step are repeated until there is no data determined to be an outlier in the outlier determination step.

[0050] [3] The outlier removal method according to [2], wherein in the outlier determination step, data whose index value is equal to or greater than a preset reference value is determined to be an outlier, and in the outlier removal step, when there are multiple data determined to be outliers in the outlier determination step, only the data that has the largest proportion of data exceeding the reference value in the distribution of prediction errors is removed.

[0051] [4] The outlier removal method according to any one of [1] to [3], wherein the index value is a median of a distribution of the prediction errors.

[0052] [5] The outlier removal method according to any one of [1] to [4], wherein at least one of mean error (ME), mean absolute error (MAE), and mean squared error (RMSE), and one of mean percent error (MPE), mean absolute error (MAPE), and mean squared error (RMSPE) are used as the prediction errors, and in the outlier determination step, data for which the index value of any one of the mean error (ME), mean absolute error (MAE), and mean squared error (RMSE) is equal to or greater than a predetermined first determination criterion value, and the index value of any one of the mean percent error (MPE), mean absolute error (MAPE), and mean squared error (RMSPE) is equal to or greater than a predetermined second determination criterion value, is determined to be an outlier.

[0053] [6] An apparatus for removing outliers included in learning data (31) including data on explanatory variables and objective variables used in machine learning, the outlier removal apparatus (1) comprising: a prediction error calculation processing unit (22) that divides the learning data (31) into teacher data and test data, creates a regression model showing a correlation between the explanatory variables and the objective variable using the teacher data, and repeats a predetermined number of times a calculation of a prediction error for the created regression model using the test data; a distribution calculation processing unit (23) that extracts, from the prediction errors obtained by the prediction error calculation processing unit (22), a prediction error for each data item included in the learning data (31) when the data item is used as the test data, and obtains an index value characterizing a distribution of the extracted prediction errors for each data item included in the learning data (31); an outlier determination processing unit (24) that determines whether each data item is an outlier based on the index value for each data item obtained by the distribution calculation processing unit (23); and an outlier removal processing unit (25) that removes data determined to be an outlier by the outlier determination processing unit (24).

[0054] (Additional Note) Although the embodiment of the present invention has been described above, the invention according to the claims is not limited to the embodiment described above. It should be noted that not all of the combinations of features described in the embodiment are essential to the means for solving the problems of the invention. The present invention can be modified appropriately without departing from the spirit of the invention. [Explanation of symbols]

[0055] 1. Outlier removal device 2...Control section 21...Data acquisition processing section 22...Prediction error calculation processing unit 23...Distribution calculation processing unit 24...Outlier determination processing unit 25…Outlier removal processing section 3...Storage section 31…Learning data 32…Forecast error data 33…Distribution data 34…Outlier candidate data 35…Outlier data

Claims

1. A method for removing outliers contained in learning data including explanatory variable and objective variable data used in machine learning, comprising: a prediction error calculation step of dividing the learning data into teacher data and test data, creating a regression model showing a correlation between the explanatory variables and the objective variable using the teacher data, and calculating a prediction error for the created regression model using the test data a predetermined number of times; a distribution calculation step of extracting a prediction error for each data item included in the learning data from the prediction errors obtained in the prediction error calculation step when the data item is used as the test data, and calculating an index value characterizing a distribution of the extracted prediction errors for each data item included in the learning data; an outlier determination step of determining whether each data item is an outlier based on the index value of each data item calculated in the distribution calculation step; An outlier removal step of removing data determined to be an outlier in the outlier determination step, Outlier removal methods.

2. In the outlier removal step, when there are a plurality of data determined to be outliers in the outlier determination step, only one of the data is removed; repeating the prediction error calculation step, the outlier determination step, and the outlier removal step until no data is determined to be an outlier in the outlier determination step; The method for removing outliers according to claim 1 .

3. In the outlier determination step, data whose index value is equal to or greater than a preset reference value is determined to be an outlier; In the outlier removal step, when there are a plurality of data determined to be outliers in the outlier determination step, only the data having the largest proportion exceeding the determination reference value in the distribution of the prediction errors is removed from the plurality of data determined to be outliers in the outlier determination step. The method for removing outliers according to claim 2 .

4. The index value is the median of the distribution of the prediction errors. The method for removing outliers according to claim 1 .

5. As the prediction error, at least one of a mean error (ME), a mean absolute error (MAE), and a mean square error (RMSE), and one of a mean percent error (MPE), a mean absolute error (MAPE), and a mean squared percent error (RMSPE) are used; In the outlier determination step, data in which any one of the index values ​​of the mean error (ME), the mean absolute error (MAE), and the mean squared error (RMSE) is equal to or greater than a first predetermined reference value, and any one of the index values ​​of the mean percent error (MPE), the mean absolute error (MAPE), and the mean squared error rate (RMSPE) is equal to or greater than a second predetermined reference value, is determined to be an outlier. The method for removing outliers according to claim 1 .

6. An apparatus for removing outliers contained in learning data including explanatory variable and objective variable data used in machine learning, comprising: a prediction error calculation processing unit that divides the learning data into teacher data and test data, creates a regression model showing a correlation between the explanatory variables and the objective variables using the teacher data, and calculates a prediction error for the created regression model using the test data, repeating this process a predetermined number of times; a distribution calculation processing unit that extracts a prediction error for each data item included in the learning data from the prediction errors obtained by the prediction error calculation processing unit when the data item is used as the test data, and calculates an index value that characterizes a distribution of the extracted prediction errors for each data item included in the learning data; an outlier determination processing unit that determines whether each data item is an outlier based on the index value of each data item calculated by the distribution calculation processing unit; and an outlier removal processing unit that removes data determined to be an outlier by the outlier determination processing unit. Outlier removal device.

Images