Training data generation program, device, and method

Abstract

To generate training data after data expansion for fairness correction that is based upon deviation of features of a training data set before expansion.SOLUTION: A training data generation device is configured to: classify minority data and majority data which are included in a training data set and associated with sensitivity attributes into similar groups based upon feature quantities; calculate an expansion number of each similar group for expansion up to the total number of majority groups while maintaining the ratio of the number of minority group in every similar group; select, for each similar group, data as many as the calculated expansion number, as data to be used for expansion, from majority data in the same similar group; and convert the selected majority data into data having features of attributes of the minority groups.SELECTED DRAWING: Figure 5

Description

[Technical Field]

[0001] The disclosed technology relates to a training data generation program, a training data generation device, and a training data generation method. [Background technology]

[0002] Recent machine learning methods often require large amounts of data as training data for training machine learning models. However, it is often difficult to collect a sufficient amount of training data. Therefore, a method is used to expand the amount of training data by converting the prepared training data to generate new training data.

[0003] For example, a data generation device has been proposed that generates training data that enables the construction of an analytical model with high generalization performance. When classifying first training data into predetermined categories using a trained analytical model, this device detects characteristic features from the first training data that contribute to classification into the predetermined categories, and generates second training data by processing the first training data according to the characteristic features.

[0004] Furthermore, for example, a neural network training device has been proposed that extracts features from training data using a neural network under training, and generates adversarial features from the extracted features using the neural network under training. This device calculates the recognition results of the neural network using the training data and the adversarial features, and trains the neural network so that the recognition results approach the desired output.

[0005] For example, a system has been proposed for expanding minority class training samples in a machine learning model that uses imbalanced training samples. The system selects training sample values ​​from a set of training samples, selects combination proportion values ​​from a continuous probability distribution, modifies the selected training sample values ​​using the combination proportion values, and generates synthetic training samples by combining the modified training sample values.

[0006] Also, for example, to balance an imbalanced training dataset that includes a majority data class and a minority data class, a system has been proposed that generates a set of data samples for the minority data class. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] International Publication No. 2021 / 130995 [Patent Document 2] International Publication No. 2018-167900 [Patent Document 3] U.S. Patent Application Publication No. 2021 / 0073671 [Patent Document 4] U.S. Patent Application Publication No. 2015 / 0088791 Summary of the Invention [Problem to be solved by the invention]

[0008] To correct the fairness of training data, it is possible to expand the training data of minority groups based on the training data of majority groups. However, if the feature biases of the training data of minority groups and the training data of majority groups are different, this may destroy the original feature bias of minority groups. Furthermore, if the original feature bias of minority groups is not maintained after expansion, the predictive accuracy of the machine learning model for minority groups may decrease.

[0009] In one aspect, the disclosed technology aims to generate a training dataset after data augmentation for fairness correction, the augmented training data being based on the bias of features in the training dataset before augmentation. [Means for solving the problem]

[0010] In one embodiment, the disclosed technology classifies a first plurality of training data having a first attribute and a second plurality of training data having a second attribute contained in a plurality of training data based on features. The disclosed technology compares the number of training data classified into a first group among the first plurality of training data with the number of training data classified into a second group. Based on this comparison, the disclosed technology selects a third plurality of training data from among the second plurality of training data classified into a third group corresponding to the first group and a fourth group corresponding to the second group. The disclosed technology converts each of the third plurality of training data into a fourth plurality of training data having the first attribute. [Effects of the Invention]

[0011] As one aspect, the training dataset after data augmentation for fairness correction has the effect of generating augmented training data based on the bias of the features of the training dataset before augmentation. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 2 is a functional block diagram of a training data generation device. [Figure 2] FIG. 10 is a diagram for explaining a first reference example. [Figure 3] FIG. 10 is a diagram for explaining a second reference example. [Figure 4] FIG. 1 is a diagram for explaining the focus of the present embodiment. [Figure 5] FIG. 2 is a diagram for explaining an outline of the processing of the present embodiment. [Figure 6] FIG. 10 is a diagram for explaining calculation of the ratio of the number of cases for each similar group. [Figure 7] FIG. 10 is a diagram for explaining calculation of the number of extensions for each similarity group. [Figure 8] FIG. 10 is a diagram for explaining selection of majority data according to the number of extensions. [Figure 9]FIG. 10 is a diagram for explaining data conversion. [Figure 10] FIG. 10 is a diagram for explaining a determination of whether or not converted data is to be adopted as extended data. [Figure 11] FIG. 1 is a block diagram showing a schematic configuration of a computer that functions as a training data generation device. [Figure 12] 10 is a flowchart illustrating an example of a training data generation process. [Figure 13] FIG. 10 is a diagram illustrating another example of calculation of the extension number for each similarity group. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, an example of an embodiment of the disclosed technology will be described with reference to the drawings.

[0014] As shown in Fig. 1, a training dataset used for machine learning training is input to a training data generation device 10 according to this embodiment. The training data generation device 10 converts the training data included in the training dataset to generate new training data (hereinafter also referred to as "augmented data"). The training data generation device 10 outputs an extended dataset obtained by adding the generated extended data to the input training dataset.

[0015] Here, when training data is grouped based on one of the attributes possessed by each training data in the training dataset, there may be a bias in the data size between groups, i.e., the number of training data included in each group. If a machine learning model is trained using a training dataset in which the number of training data is biased due to an attribute that requires consideration of fairness (hereinafter referred to as a "sensitive attribute"), there is a high possibility that the prediction results of the machine learning model will be discriminatory. Therefore, it is desirable to correct such bias in the data size in the training dataset. Hereinafter, in grouping based on sensitive attributes, the group with a larger data size will be referred to as the "majority group," and the data included in the majority group will be referred to as the "majority data." Furthermore, the group with a smaller data size will be referred to as the "minority group," and the data included in the minority group will be referred to as the "minority data."

[0016] Furthermore, when the size of the training data is unevenly distributed among groups classified by sensitive attributes, the small amount of training data for minority groups tends to reduce the prediction accuracy of machine learning models for minority data compared to majority data. Therefore, it is necessary to improve the prediction accuracy for minority groups by expanding the data.

[0017] As a method for expanding data for minority groups, the following first reference example can be considered. For example, as shown in FIG. 2, assume that the training data are facial images of people. In this case, when the training data is grouped by the sensitive attribute "gender (male, female)," the group of male gender (hereinafter referred to as the "male group") is the minority group. Also, assume that the group of female gender (hereinafter referred to as the "female group") is the majority group. In the first reference example, facial images of the training data for the female group (hereinafter also referred to as "female data"), which is the majority data, are converted into masculine facial images, and these are used as training data for the male group (hereinafter also referred to as "male data"), which is the minority data. In this way, the first reference example expands the minority group and equalizes the data scale of the majority group and the minority group. However, the first reference example has a problem in that if the characteristics for the task to be predicted differ between the groups, data having those characteristics cannot be expanded. For example, as shown in the example of FIG. 2, in the converted facial image, the facial features may be converted to have masculine features, but the hairstyle may still have feminine features. In such a case, for example, in a task where long hair is a positive example for women and short hair is a positive example for men, the prediction accuracy for the male group will not improve.

[0018] Furthermore, the following second reference example can be considered as a method for expanding data for minority groups. In the second reference example, the minority group is expanded by generating new data by performing processes such as rotation, enlargement, reduction, and color change on the minority data, as shown in Fig. 3. However, in the second reference example, since the amount of training data for the same case increases, the training data for the expanded minority group lacks diversity in features, which makes it susceptible to overfitting.

[0019] Therefore, this embodiment proposes a data expansion method that allows for diverse expressions while maintaining the characteristics of minority data. This embodiment focuses on the fact that the majority group may contain training data with characteristics similar to those of the minority data. For example, as shown in FIG. 4, assume that minority data, i.e., male data, has a characteristic that there are more short-haired individuals than long-haired individuals. In this case, the training data generation device 10 according to this embodiment preferentially selects facial images with short hair from the majority data, i.e., female data, as targets for expansion. The training data generation device 10 also suppresses deviations between the distribution of the label to be predicted by the machine learning model and the characteristics of the training data. For example, if training data for straight and short hair in a male group accounts for 30% of all training data for the male group, the training data generation device 10 performs data expansion so that this proportion remains unchanged after expansion.

[0020] The functional units of the training data generation device 10 according to this embodiment will be described in detail below. To explain a specific example in this embodiment, a task is assumed in which a person's facial image is predicted to be "attractive" or "unattractive." The sensitive attribute is gender, and it is considered fair if there is little difference in prediction accuracy between men and women. Furthermore, this embodiment assumes that the number of training data for the male group is insufficient and is even smaller than the number of training data for the female group. In other words, it is assumed that the male data is minority data and the female data is majority data. In this task, it is assumed that hairstyle characteristics have a strong influence on whether a person is "attractive."

[0021] 1, the training data generation device 10 functionally includes a control unit 12. More specifically, the control unit 12 includes a classification unit 14, a selection unit 16, a conversion unit 18, and a determination unit 20.

[0022] The classification unit 14 classifies each of a first plurality of training data having a first attribute and a second plurality of training data having a second attribute included in the training data set based on the feature amount. Here, the first attribute is male and the second attribute is female. That is, the first plurality of training data is training data classified into a male group, which is a minority group, and the second plurality of training data is training data classified into a female group, which is a majority group.

[0023] Specifically, the classification unit 14 extracts features from each training data. For example, when each training data is input to a deep neural network, which is an example of a machine learning model, the classification unit 14 extracts values ​​output from at least one of an intermediate layer and an output layer of the deep neural network as features of the training data. Then, as shown in the upper diagram of FIG. 5, the classification unit 14 clusters the training data based on the similarity of the features. In the example of FIG. 5, black circles represent male data and crosses represent female data. Hereinafter, groups classified by clustering based on the similarity of the features are referred to as "similar groups." In the example of FIG. 5, the training data is classified into one of four similar groups, A, B, C, and D. When the above-mentioned features are used, there is a tendency for groups to be classified according to hairstyle characteristics, such as similar group A being short ∧ straight hair, similar group B being short ∧ curly hair, similar group C being long ∧ curly hair, and similar group D being long ∧ straight hair.

[0024] The selection unit 16 compares the number of training data items classified into the first similarity group among the minority data with the number of training data items classified into the second similarity group. Based on this comparison, the selection unit 16 selects training data to be used for extension from the training data items classified into the third similarity group corresponding to the first similarity group among the majority data. The selection unit 16 also selects training data to be used for extension from the training data items classified into the fourth similarity group corresponding to the second similarity group among the training data items in the majority group. The training data to be used for extension is an example of the "third plurality of training data" of the disclosed technology. In addition, in this embodiment, the first similarity group and the third similarity group are the same, and the second similarity group and the fourth similarity group are the same.

[0025] Specifically, as shown in FIG. 6, the selection unit 16 counts the number of minority data (male data) classified into each similar group and calculates the proportion of each similar group. The proportion is the number of minority data classified into each similar group divided by the total number of minority data (data size of the minority group). The selection unit 16 also counts the total number of majority data (female data) (data size of the majority group). In the example of FIG. 6, for reference, the number and proportion of each similar group are also shown for the female group, which is the majority group.

[0026] As shown in FIG. 7, the selection unit 16 calculates an expansion number for each similar group for the minority group so as to equalize the data size of the minority group and the data size of the majority group while maintaining the calculated ratio. Note that FIG. 7 illustrates a case in which the expansion number is calculated so that the expanded minority group and the majority group have the same number of items. However, making the data size equal includes not only the case in which the minority group and the majority group have the same number of data items, but also the case in which the difference between the data size of the minority group and the data size of the majority group is within a first threshold. The first threshold is a value that is determined in advance so that the data size of the minority group and the data size of the majority group are considered to be equal. For each similar group, the selection unit 16 selects, from each corresponding similar group, a number of majority data items corresponding to the calculated expansion number as training data to be used for expansion.

[0027] More specifically, as shown in the left diagram of FIG. 8, when the number of majority data items classified into the same similarity group is greater than the calculated extension number for that similarity group, the selection unit 16 selects the majority data items from that similarity group by the calculated extension number. The example in the left diagram of FIG. 8 shows a case where the number of majority data items classified into the same similarity group is four and the calculated extension number is two. In this case, the selection unit 16 selects two majority data items from that similarity group. In this case, the selection unit 16 selects the majority data items from the majority data items classified into the same similarity group in descending order of similarity with the minority data items classified into that similarity group by the extension number. The selection unit 16 may use, for example, the distance between the center of the minority data items in that similarity group and each of the majority data items as the similarity between the majority data items and the minority data items.

[0028] Furthermore, as shown in the right diagram of Fig. 8, when the number of majority data classified into the same similar group is equal to or less than the extension number calculated for that similar group, the selection unit 16 selects all of the majority data in that similar group. The example in the right diagram of Fig. 8 shows a case where the number of majority data classified into the same similar group is 3 and the calculated extension number is 4. In this case, the selection unit 16 selects three majority data from that similar group.

[0029] The conversion unit 18 converts each of the majority data selected by the selection unit 16, which is training data to be used for extension, into data having the characteristics of a first attribute, i.e., the attribute of a minority group. Specifically, as shown in FIG. 9, the conversion unit 18 converts female data (facial image) which is the majority data into a facial image having male characteristics. For example, the conversion unit 18 performs image conversion using a machine-learned generative model such as GAN (Generative Adversarial Networks). The data converted by the conversion unit 18 is an example of the "fourth plurality of training data" of the disclosed technology.

[0030] The conversion unit 18 can perform expansion taking into consideration the characteristics of the minority data because it uses majority data of the same similar group for expansion, as shown in the middle diagram of Fig. 5. Furthermore, the conversion unit 18 expands the minority group so that the data scale is equivalent to that of the majority data while maintaining the proportion of the number of records in each original similar group, as shown in the bottom diagram of Fig. 5, so it is possible to maintain the bias in the characteristics of the minority group.

[0031] The determination unit 20 determines whether or not to adopt the data converted by the conversion unit 18 as extended data. Specifically, as shown in FIG. 10 , if the converted data is classified into the same similarity group as the similarity group of the majority data before conversion, the determination unit 20 adopts the converted data as extended data. On the other hand, if the converted data is not classified into the same similarity group as the similarity group of the majority data before conversion, the determination unit 20 does not adopt the converted data as extended data. For example, the conversion unit 18 determines whether or not the converted data is classified into the same similarity group by applying the converted data to a classification model generated during clustering by the classification unit 14. Furthermore, the conversion unit 18 may determine whether or not the converted data is classified into the same similarity group based on, for example, whether or not the distance between the cluster center of the same similarity group and the converted data is equal to or greater than a predetermined value.

[0032] The determination unit 20 excludes converted data determined not to be adopted as extended data from the converted data group, and sets the remaining data as extended data. Note that, as shown in the left diagram of FIG. 8, when the selection unit 16 selects majority data, unselected majority data may exist. In this case, the determination unit 20 may cause the selection unit 16 to reselect majority data other than the majority data that was the basis for the data determined not to be adopted. The determination unit 20 adds the extended data to the original training dataset to set up an extended dataset, and outputs the extended dataset.

[0033] The training data generation device 10 may be realized, for example, by a computer 40 shown in FIG. 11 . The computer 40 includes a CPU (Central Processing Unit) 41, a memory 42 serving as a temporary storage area, and a non-volatile storage device 43. The computer 40 also includes an input / output device 44 such as an input device and a display device, and an R / W (Read / Write) device 45 that controls reading and writing of data from and to a storage medium 49. The computer 40 also includes a communication I / F (Interface) 46 that is connected to a network such as the Internet. The CPU 41, memory 42, storage device 43, input / output device 44, R / W device 45, and communication I / F 46 are connected to one another via a bus 47.

[0034] The storage device 43 is, for example, a hard disk drive (HDD), a solid state drive (SSD), a flash memory, etc. The storage device 43 serving as a storage medium stores a training data generation program 50 for causing the computer 40 to function as the training data generation device 10. The training data generation program 50 includes a classification process control instruction 54, a selection process control instruction 56, a conversion process control instruction 58, and a determination process control instruction 60.

[0035] The CPU 41 reads the training data generation program 50 from the storage device 43, loads it into the memory 42, and sequentially executes the control instructions of the training data generation program 50. The CPU 41 operates as the classification unit 14 shown in FIG. 1 by executing the classification process control instruction 54. The CPU 41 operates as the selection unit 16 shown in FIG. 1 by executing the selection process control instruction 56. The CPU 41 operates as the conversion unit 18 shown in FIG. 1 by executing the conversion process control instruction 58. The CPU 41 operates as the judgment unit 20 shown in FIG. 1 by executing the judgment process control instruction 60. As a result, the computer 40 that has executed the training data generation program 50 functions as the training data generation device 10. The CPU 41 that executes the program is hardware.

[0036] The functions realized by the training data generation program 50 may be realized by, for example, a semiconductor integrated circuit, more specifically, an ASIC (Application Specific Integrated Circuit).

[0037] Next, the operation of the training data generation device 10 according to this embodiment will be described. When a training data set is input to the training data generation device 10 and an instruction to generate augmented data is given, the training data generation device 10 executes the training data generation process shown in Fig. 12. Note that the training data generation process is an example of a training data generation method of the disclosed technology.

[0038] In step S10, the classification unit 14 acquires the training data set input to the training data generation device 10. Then, the classification unit 14 extracts features from each training data, and classifies the training data into one of the similarity groups by clustering the training data based on the similarity of the features.

[0039] Next, in step S12, the selection unit 16 counts the number of minority data items classified into each similarity group and calculates the proportion of the number of items for each similarity group. The selection unit 16 also counts the total number of majority data items. Next, in step S14, the selection unit 16 calculates an expansion number for each similarity group to make the data scale of the minority group equal to the data scale of the majority group while maintaining the calculated proportion for the minority group.

[0040] Next, in step S16, the selection unit 16 determines, for each similarity group, whether the number of majority data items classified into the same similarity group is greater than the expansion number calculated for that similarity group. If the number of majority data items is greater, the process proceeds to step S18. If the expansion number is equal to or greater than the number of minority data items, the process proceeds to step S20.

[0041] In step S18, the selection unit 16 selects the number of majority data to be expanded from the majority data classified into the same similarity group in descending order of similarity to the minority data classified into the same similarity group. Meanwhile, in step S20, the selection unit 16 selects all majority data in the same similarity group. Note that the processes of steps S16 to S20 are performed for each similarity group.

[0042] Next, in step S22, the conversion unit 18 converts each of the majority data selected in step S18 or S20 into data having the attribute characteristics of the minority group. Next, in step S24, if the converted data is not classified into the same similar group as the similar group of the majority data before conversion, the determination unit 20 excludes the converted data from the converted data group and outputs the remaining data as an extended data set. Then, the training data generation process ends.

[0043] As described above, the training data generation device according to this embodiment classifies minority data and majority data related to sensitive attributes included in a training dataset into similar groups based on their features. The training data generation device also calculates an extension number for each similar group to expand the training dataset to the total number of majority groups while maintaining the proportion of minority groups in each similar group. For each similar group, the training data generation device selects, as data to be used for extension, the calculated number of pieces of data from the majority data of the same similar group. The training data generation device then converts the selected majority data into data having the attributes of the minority group to generate extended data. This allows for the generation of extended training data so that the training dataset after data extension for fairness correction maintains the bias of the features of the training dataset before extension. Furthermore, because the extended data is generated by converting majority data similar to the features of the minority data, the prediction accuracy of minority groups is improved.

[0044] Note that, since there are various concepts and standards for fairness, there are also various fairness indices. In the above embodiment, an accuracy parity index is assumed, which defines fairness as matching prediction accuracy. Therefore, in the above embodiment, data expansion is performed to equalize the data scale between groups of sensitive attributes. Another representative fairness index is, for example, the demographic parity index. This index defines fairness as matching the positive example rate between groups of sensitive attributes. In this case, as shown in FIG. 13 , the selection unit selects training data to be used for expansion from the majority group of the same similar group so that the positive example rate for each similar group of the minority group is similar to that of the majority group of the same similar group.

[0045] Specifically, as in the above embodiment, the selection unit calculates an expansion number for each similar group so that the data scales of the minority group and the majority group are equivalent while maintaining the ratio of the number of cases per similar group for the minority group. Then, when selecting majority data from the same similar group for each similar group, the selection unit selects the majority data so that the positive example rates of the similar groups are equivalent. Note that "equalizing the positive example rates" includes not only cases where the positive example rates of the minority group and the majority group are the same, but also cases where the difference between the positive example rate of the minority group and the positive example rate of the majority group is within a second or third threshold. The second and third thresholds are thresholds for each similar group, and are predetermined values ​​at which the positive example rates of the minority group and the majority group are considered equivalent. More specifically, when the positive example rate of the minority group is lower than that of the majority group, the selection unit preferentially selects the majority data of the positive examples. On the other hand, when the positive example rate of the minority group is higher than the positive example rate of the majority group, the selection unit preferentially selects the negative example majority data.

[0046] In the above embodiment, the training data generation program is stored (installed) in advance in a storage device, but this is not limiting. The program according to the disclosed technology may be provided in a form stored in a storage medium such as a CD-ROM, a DVD-ROM, or a USB memory.

[0047] The following additional notes are provided regarding the above-described embodiments.

[0048] (Appendix 1) classifying a first plurality of training data having a first attribute and a second plurality of training data having a second attribute, the first plurality of training data being included in the plurality of training data, based on the feature amount; selecting a third plurality of training data from among the second plurality of training data classified into a third group corresponding to the first group and a fourth group corresponding to the second group, based on a comparison between the number of training data classified into a first group and the number of training data classified into a second group among the first plurality of training data; converting each of the third plurality of training data into a fourth plurality of training data having the first attribute; A training data generation program that causes a computer to execute a process.

[0049] (Appendix 2) the process of selecting the third plurality of training data includes selecting, from each of the third group and the fourth group, a number of training data corresponding to an expansion number when expanding the number of training data in each of the first group and the second group so that a difference between a total number of training data classified into the first group and the second group and a total number of training data classified into the third group and the fourth group among the second plurality of training data is within a first threshold while maintaining a ratio between a number of training data classified into the first group among the first plurality of training data and a number of training data classified into the second group. 1. The training data generation program described in Appendix 1.

[0050] (Appendix 3) The process of selecting the third plurality of training data includes: selecting the training data classified into the third group so that a difference between a positive example rate of the training data classified into the first group and a positive example rate of the training data classified into the third group is within a second threshold; selecting the training data classified into the fourth group so that a difference between a positive example rate of the training data classified into the second group and a positive example rate of the training data classified into the fourth group is within a third threshold. Attachment 2. A training data generation program.

[0051] (Appendix 4) The number corresponding to the extension number is When the number of training data classified into the third group is greater than the extension number for the first group, the extension number is set to the first group, and when the number of training data classified into the third group is equal to or less than the extension number for the first group, the extension number is set to the number of training data classified into the third group; If the number of training data classified into the fourth group is greater than the extension number for the second group, the extension number is set to the second group, and if the number of training data classified into the fourth group is equal to or less than the extension number for the second group, the extension number is set to the number of training data classified into the fourth group. 4. The training data generation program according to claim 2 or 3.

[0052] (Appendix 5) If the number of training data classified into the third group is greater than the number of extensions for the first group, select training data from the training data classified into the third group in descending order of similarity to the training data classified into the first group, the number of extensions for the first group; If the number of training data classified into the fourth group is greater than the number of extensions for the second group, select training data of the number of extensions for the second group from the training data classified into the fourth group in descending order of similarity to the training data classified into the second group. Attachment 4. A training data generation program.

[0053] (Appendix 6) excluding, from the fourth plurality of training data obtained by converting the third plurality of training data selected from the third group, training data that does not fall into the first group; excluding, from the fourth plurality of training data obtained by converting the third plurality of training data selected from the fourth group, training data that is not classified into the second group; 6. The training data generation program according to claim 1, for causing the computer to execute a process including the steps of:

[0054] (Appendix 7) classifying a first plurality of training data having a first attribute and a second plurality of training data having a second attribute, the first plurality of training data being included in the plurality of training data, based on the feature amount; selecting a third plurality of training data from among the second plurality of training data classified into a third group corresponding to the first group and a fourth group corresponding to the second group, based on a comparison between the number of training data classified into a first group and the number of training data classified into a second group among the first plurality of training data; converting each of the third plurality of training data into a fourth plurality of training data having the first attribute; A training data generation device comprising a control unit that executes processing.

[0055] (Appendix 8) the process of selecting the third plurality of training data includes selecting, from each of the third group and the fourth group, a number of training data corresponding to an expansion number when expanding the number of training data in each of the first group and the second group so that a difference between a total number of training data classified into the first group and the second group and a total number of training data classified into the third group and the fourth group among the second plurality of training data is within a first threshold while maintaining a ratio between a number of training data classified into the first group among the first plurality of training data and a number of training data classified into the second group. 8. The training data generation device according to claim 7.

[0056] (Appendix 9) The process of selecting the third plurality of training data includes: selecting the training data classified into the third group so that a difference between a positive example rate of the training data classified into the first group and a positive example rate of the training data classified into the third group is within a second threshold; selecting the training data classified into the fourth group so that a difference between a positive example rate of the training data classified into the second group and a positive example rate of the training data classified into the fourth group is within a third threshold. 9. The training data generation device according to claim 8.

[0057] (Appendix 10) The number corresponding to the extension number is When the number of training data classified into the third group is greater than the extension number for the first group, the extension number is set to the first group, and when the number of training data classified into the third group is equal to or less than the extension number for the first group, the extension number is set to the number of training data classified into the third group; If the number of training data classified into the fourth group is greater than the extension number for the second group, the extension number is set to the second group, and if the number of training data classified into the fourth group is equal to or less than the extension number for the second group, the extension number is set to the number of training data classified into the fourth group. 10. The training data generation device according to claim 8 or 9.

[0058] (Appendix 11) If the number of training data classified into the third group is greater than the number of extensions for the first group, select training data from the training data classified into the third group in descending order of similarity to the training data classified into the first group, the number of extensions for the first group; If the number of training data classified into the fourth group is greater than the number of extensions for the second group, select training data of the number of extensions for the second group from the training data classified into the fourth group in descending order of similarity to the training data classified into the second group. 11. The training data generation device according to claim 10.

[0059] (Appendix 12) excluding, from the fourth plurality of training data obtained by converting the third plurality of training data selected from the third group, training data that does not fall into the first group; excluding, from the fourth plurality of training data obtained by converting the third plurality of training data selected from the fourth group, training data that is not classified into the second group; 12. The training data generation device according to claim 7, wherein the control unit executes the processing.

[0060] (Appendix 13) classifying a first plurality of training data having a first attribute and a second plurality of training data having a second attribute, the first plurality of training data being included in the plurality of training data, based on the feature amount; selecting a third plurality of training data from among the second plurality of training data classified into a third group corresponding to the first group and a fourth group corresponding to the second group, based on a comparison between the number of training data classified into a first group and the number of training data classified into a second group among the first plurality of training data; converting each of the third plurality of training data into a fourth plurality of training data having the first attribute; A training data generation method characterized in that processing is executed by a computer.

[0061] (Appendix 14) the process of selecting the third plurality of training data includes selecting, from each of the third group and the fourth group, a number of training data corresponding to an expansion number when expanding the number of training data in each of the first group and the second group so that a difference between a total number of training data classified into the first group and the second group and a total number of training data classified into the third group and the fourth group among the second plurality of training data is within a first threshold while maintaining a ratio between a number of training data classified into the first group among the first plurality of training data and a number of training data classified into the second group. 14. The training data generation method of claim 13.

[0062] (Appendix 15) The process of selecting the third plurality of training data includes: selecting the training data classified into the third group so that a difference between a positive example rate of the training data classified into the first group and a positive example rate of the training data classified into the third group is within a second threshold; selecting the training data classified into the fourth group so that a difference between a positive example rate of the training data classified into the second group and a positive example rate of the training data classified into the fourth group is within a third threshold. 15. The training data generation method of claim 14.

[0063] (Appendix 16) The number corresponding to the extension number is When the number of training data classified into the third group is greater than the extension number for the first group, the extension number is set to the first group, and when the number of training data classified into the third group is equal to or less than the extension number for the first group, the extension number is set to the number of training data classified into the third group; If the number of training data classified into the fourth group is greater than the extension number for the second group, the extension number is set to the second group, and if the number of training data classified into the fourth group is equal to or less than the extension number for the second group, the extension number is set to the number of training data classified into the fourth group. 16. The training data generation method according to claim 14 or 15.

[0064] (Appendix 17) If the number of training data classified into the third group is greater than the number of extensions for the first group, select training data from the training data classified into the third group in descending order of similarity to the training data classified into the first group, the number of extensions for the first group; If the number of training data classified into the fourth group is greater than the number of extensions for the second group, select training data of the number of extensions for the second group from the training data classified into the fourth group in descending order of similarity to the training data classified into the second group. 17. The training data generation method of claim 16.

[0065] (Appendix 18) excluding, from the fourth plurality of training data obtained by converting the third plurality of training data selected from the third group, training data that does not fall into the first group; excluding, from the fourth plurality of training data obtained by converting the third plurality of training data selected from the fourth group, training data that is not classified into the second group; 18. The training data generation method according to any one of Supplementary Note 13 to Supplementary Note 17, wherein the computer executes a process including:

[0066] (Appendix 19) classifying a first plurality of training data having a first attribute and a second plurality of training data having a second attribute, the first plurality of training data being included in the plurality of training data, based on the feature amount; selecting a third plurality of training data from among the second plurality of training data classified into a third group corresponding to the first group and a fourth group corresponding to the second group, based on a comparison between the number of training data classified into a first group and the number of training data classified into a second group among the first plurality of training data; converting each of the third plurality of training data into a fourth plurality of training data having the first attribute; A non-transitory storage medium storing a training data generation program that causes a computer to execute a process. [Explanation of symbols]

[0067] 10 Training data generator 12 Control Unit 14 Classification section 16 Selection section 18 Conversion unit 20 Judgment section 40 Computer 41 CPU 42 memory 43 Storage device 44 Input / Output Devices 46 Communication I / F 45 R / W device 47 Bus 49 Storage medium 50 Training data generation program 54 Classification Process Control Instructions 56 Selective Process Control Instructions 58 Conversion Process Control Instructions 60 Decision Process Control Instructions

Claims

1. classifying a first plurality of training data having a first attribute and a second plurality of training data having a second attribute, the first plurality of training data being included in the plurality of training data, based on the feature amount; selecting a third plurality of training data from among the second plurality of training data classified into a third group corresponding to the first group and a fourth group corresponding to the second group, based on a comparison between the number of training data classified into a first group and the number of training data classified into a second group among the first plurality of training data; converting each of the third plurality of training data into a fourth plurality of training data having the first attribute; A training data generation program that causes a computer to execute a process.

2. the process of selecting the third plurality of training data includes selecting, from each of the third group and the fourth group, a number of training data corresponding to an expansion number when the number of training data in each of the first group and the second group is expanded so that a difference between a total number of training data classified into the first group and the second group and a total number of training data classified into the third group and the fourth group among the second plurality of training data is within a first threshold while maintaining a ratio between a number of training data classified into the first group among the first plurality of training data and a number of training data classified into the second group. The training data generation program according to claim 1 .

3. The process of selecting the third plurality of training data includes: selecting the training data classified into the third group so that a difference between a positive example rate of the training data classified into the first group and a positive example rate of the training data classified into the third group is within a second threshold; selecting the training data classified into the fourth group such that a difference between a positive example rate of the training data classified into the second group and a positive example rate of the training data classified into the fourth group is within a third threshold. The training data generation program according to claim 2 .

4. The number corresponding to the extension number is When the number of training data classified into the third group is greater than the extension number for the first group, the extension number is set to the first group, and when the number of training data classified into the third group is equal to or less than the extension number for the first group, the extension number is set to the number of training data classified into the third group; If the number of training data classified into the fourth group is greater than the extension number for the second group, the extension number is set to the second group, and if the number of training data classified into the fourth group is equal to or less than the extension number for the second group, the extension number is set to the number of training data classified into the fourth group.

4. The training data generation program according to claim 2 or 3.

5. If the number of training data classified into the third group is greater than the number of extensions for the first group, select training data from the training data classified into the third group in descending order of similarity to the training data classified into the first group, the number of extensions for the first group; when the number of training data classified into the fourth group is greater than the number of extensions for the second group, selecting training data, the number of which corresponds to the number of extensions for the second group, from the training data classified into the fourth group in descending order of similarity to the training data classified into the second group; The training data generation program according to claim 4.

6. excluding, from the fourth plurality of training data obtained by converting the third plurality of training data selected from the third group, training data that does not fall into the first group; excluding, from the fourth plurality of training data obtained by converting the third plurality of training data selected from the fourth group, training data that is not classified into the second group; The training data generation program according to any one of claims 1 to 3, for causing the computer to execute a process including:

7. classifying a first plurality of training data having a first attribute and a second plurality of training data having a second attribute, the first plurality of training data being included in the plurality of training data, based on the feature amount; selecting a third plurality of training data from among the second plurality of training data classified into a third group corresponding to the first group and a fourth group corresponding to the second group, based on a comparison between the number of training data classified into a first group and the number of training data classified into a second group among the first plurality of training data; converting each of the third plurality of training data into a fourth plurality of training data having the first attribute; A training data generation device comprising a control unit that executes processing.

8. classifying a first plurality of training data having a first attribute and a second plurality of training data having a second attribute, the first plurality of training data being included in the plurality of training data, based on the feature amount; selecting a third plurality of training data from among the second plurality of training data classified into a third group corresponding to the first group and a fourth group corresponding to the second group, based on a comparison between the number of training data classified into a first group and the number of training data classified into a second group among the first plurality of training data; converting each of the third plurality of training data into a fourth plurality of training data having the first attribute; A training data generation method characterized in that processing is executed by a computer.

Images