Data generation device, and computer program
Patent Information
- Application Number
- JP2024051674
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-03-27
- Publication Date
- 2025-10-09
Smart Images
Figure 2025150664000001_ABST
Abstract
Description
[Technical Field]
[0001] The present disclosure relates to techniques for generating data. [Background technology]
[0002] Conventionally, a technique is known in which normal data and abnormal data are arranged in the same latent space to generate pseudo-generated data, which is abnormal data different from actual abnormal data (Patent Document 1). [Prior art documents] [Patent documents]
[0003] [Patent Document 1] International Publication No. 2021 / 161405 Summary of the Invention [Problem to be solved by the invention]
[0004] In conventional techniques, the characteristics of pseudo-generated data and the characteristics of actual anomalous data are similar, and a bias in the data distribution may occur in a data collection of pseudo-generated data and anomalous data. When a bias in the data distribution occurs, for example, when the data collection is used to train a machine learning model, a trained machine learning model with poor versatility may be generated. Furthermore, for example, in order to prevent a decrease in the versatility of the machine learning model, it may become necessary to select data to be used for training from the data collection. [Means for solving the problem]
[0005] According to a first aspect of the present disclosure, there is provided a data generation device including: a determination unit that performs a determination process to determine a generation area for generating extended coordinate data on the coordinate system based on positions on the coordinate system in a plurality of existing coordinate data obtained by converting each of a plurality of existing data into a predetermined coordinate dimension; and a data generation unit that performs a generation process to generate the extended coordinate data such that bias in data distribution within the generation area is smaller than before the extended coordinate data is generated.
[0006] According to a second aspect of the present disclosure, there is provided a computer program that causes a computer to implement: a determination function that executes a determination process to determine a generation area on the coordinate system for generating extended coordinate data based on positions on the coordinate system of a plurality of existing coordinate data obtained by converting each of a plurality of existing data into a predetermined coordinate dimension; and a data generation function that executes a generation process to generate the extended coordinate data so that the bias in data distribution within the generation area is smaller than before the extended coordinate data was generated. [Brief explanation of the drawings]
[0007] [Figure 1] FIG. 1 is a diagram showing a data generating device according to a first embodiment. [Figure 2] FIG. 1 is a diagram for explaining a generative model. [Figure 3] FIG. 10 is a diagram showing a state after existing data has been converted into existing coordinate data. [Figure 4] FIG. 10 is a diagram for explaining a predetermined distribution. [Figure 5] 10 is a flowchart showing a data generation process executed by the data generation device. [Figure 6] FIG. 1 is a diagram for explaining a data generation process. [Figure 7] FIG. 2 is a diagram for explaining the data generation process. [Figure 8] FIG. 10 is a diagram for explaining a data generation process according to the second embodiment. DETAILED DESCRIPTION OF THE INVENTION
[0008] A. First embodiment: FIG. 1 is a diagram illustrating a data generating device 10 according to a first embodiment. The data generating device 10 is a device that generates extended data DTb, which is pseudo data, using existing data DTa. The data generating device 10 is also a device that trains a machine learning model using the existing data DTa and the extended data DTb. The data generating device 10 may be a single device, or may be configured by multiple devices that can communicate data with each other.
[0009] The data generating device 10 includes a processor 20, a storage device 30, and an interface 40. The data generating device 10 is, for example, an electronic computer such as a personal computer.
[0010] The interface 40 is an input / output interface that outputs data to the output device 50 and receives data from the input device 52. The output device 50 is a device that outputs data, such as a display device such as an LCD monitor. The input device 52 is a device that can input data, such as a mouse or keyboard.
[0011] The storage device 30 is configured with memories such as RAM and ROM. The storage device 30 stores a program 32 for realizing each function of the processor 20, a trained model 33 which is a machine learning model, a generative model group 34G, and a data group DTG.
[0012] The trained model 33 is, for example, a machine learning model that uses a neural network as an algorithm. In this embodiment, the trained model 33 is a discrimination model that, when target data is input, outputs a class that is a result corresponding to the target data. Specifically, the trained model 33 is a substance discrimination model that, when inputting one-dimensional spectral data that is waveform data such as voltage or temperature as target data, outputs a class representing the type of substance corresponding to the target data. Note that the trained model 33 is not limited to this, and may also be a discrimination model or a regression model that, when inputting image data as target data, outputs whether the target data is good or bad. The trained model 33 is trained using data from a data group DTG, which will be described later.
[0013] The generative model group 34G includes multiple generative models 34 for each class of data used to train the pre-training machine learning model. The multiple generative models 34 are prepared in a number N corresponding to the type of material to be discriminated. The generative model group 34G may be composed of only one generative model 34. Each generative model 34 is a model having an algorithm for generating the extended data DTb. In this embodiment, each generative model 34 is an AAE (Adversarial Auto Encoder). The AAE can convert multiple pieces of data to have a predetermined distribution on a coordinate system representing a latent variable Z, and generate the converted data as coordinate data for the latent variable Z. In this embodiment, the latent variable Z is represented as two-dimensional coordinates on the X and Y axes. Details of the generative models 34 will be described later. In other embodiments, each generative model 34 may be another generative model, for example, a variational autoencoder.
[0014] The data group DTG includes an existing data group DTaG and an extended data group DTbG. The existing data group DTaG includes multiple existing data DTa. The multiple existing data DTa are data representing physical quantities acquired by a sensor. In this embodiment, the multiple existing data DTa are one-dimensional spectral data that detects physical quantities of a target substance over time. The multiple existing data DTa are provided for each class that represents the type of target substance. The multiple existing data DTa are used for training the corresponding generative model 34 for each class. In addition, the multiple existing data DTa for all classes are also used for training the pre-training machine learning model.
[0015] The extended data group DTbG includes a plurality of extended data DTb. Each of the plurality of extended data DTb is pseudo data generated by being input to each trained generative model 34. The plurality of extended data DTb is data in the same format as the existing data DTa, and is data representing physical quantities.
[0016] The processor 20 controls the operation of the data generating device 10 by executing a program 32 stored in the storage device 30. Specifically, by executing the program 32, the processor 20 functions as a first conversion unit 21, a second conversion unit 22, a determination unit 23, a data generation unit 24, a model generation unit 26, a model use unit 28, and an output unit 29. Note that some of the functions executed by the processor 20 may be realized by hardware circuits. Here, in this disclosure, the term "processor" encompasses a CPU and a GPU.
[0017] The first conversion unit 21 inputs each of the multiple existing data DTa into a generative model 34 corresponding to the corresponding class, and executes a first conversion process to convert the data into existing coordinate data DAa. The existing coordinate data DAa is expressed in a predetermined coordinate dimension that defines a latent space representing the latent variable Z. That is, the first conversion unit 21 executes the first conversion process for each class for the multiple existing data DTa classified into each of the multiple classes. In this embodiment, the coordinate dimension is two. Note that in other embodiments, the coordinate dimension may be one dimension or three or more dimensions.
[0018] The second conversion unit 22 executes a second conversion process to convert data expressed in a predetermined coordinate dimension, for example, extended coordinate data DAb described later, into the format of the existing data DTa to generate extended data DTb.
[0019] The determination unit 23 executes a determination process using the generative model 34 to determine a generation region Rga for generating extended coordinate data DAb on the coordinate system based on the coordinate positions of the multiple existing coordinate data DAa converted by the first conversion unit 21. The extended coordinate data DAb is data expressed in coordinate dimensions, and in this embodiment, is data expressed in coordinates of the X and Y axes. The extended coordinate data DAb is input to a generative model 34 corresponding to the corresponding class by the data generation unit 24, which will be described later, and is converted into extended data DTb. Details of the determination process will be described later.
[0020] The data generation unit 24 executes a generation process to generate the extended coordinate data DAb so that the bias of the data distribution within the generation region Rga is smaller than before the generation of the extended coordinate data DAb. Before the generation of the extended coordinate data DAb, a plurality of existing coordinate data DAa are arranged in the coordinates of each class. The details of the generation process will be described later.
[0021] The model generation unit 26 trains a pre-learning machine learning model using a plurality of existing data DTa and a plurality of extended data DTb to generate a trained model 33. The plurality of existing data DTa and the plurality of extended data DTb are data of the data group DTG, and are data of all classes.
[0022] The model use unit 28 inputs discrimination data as target data into the trained model 33 and outputs the results. The discrimination data is data in the same format as the existing data DTa and the extended data DTb, and in this embodiment, is one-dimensional spectral data. The output unit 29 transmits the results output by the model use unit 28 to the output device 50 via the interface 40. Upon receiving the results, the output device 50 outputs the results by displaying them, for example.
[0023] FIG. 2 is a diagram illustrating the generative model 34. The generative model 34 shown in FIG. 2 is a model for a specific class of material types, but other classes have a similar configuration. The generative model 34 includes an encoder unit 343 and a decoder unit 346. The encoder unit 343 performs dimensional compression on the input existing data DTa, i.e., extracts features, and represents them using latent variables Z as existing coordinate data DAa expressed in the dimensions of the coordinates to be placed. The decoder unit 346 generates and outputs data DTt, such as expanded data DTb, from the latent variables Z, which are coordinate data. The generative model 34 is trained in advance so that the similarity between the input data and the output data is equal to or greater than a predetermined reference value and so that the input data follows a predetermined distribution on the coordinates where the latent variables Z are placed. In this embodiment, the generative model 34 is trained using each existing data DTa in the existing data group DTaG. The predetermined distribution is specified by a user or the like, such as a normal distribution. In this embodiment, the type of predetermined distribution is a two-dimensional normal distribution. The distribution is specified by previously setting parameters corresponding to the mean, variance, etc. of the two-dimensional normal distribution by the user, etc. Note that if the generative model 34 is trained in advance by another device, only the decoder unit 346 may be stored in the storage device 30 as the generative model 34.
[0024] Fig. 3 is a diagram showing a state after a plurality of existing data DTa has been converted into existing coordinate data DAa so as to have a predetermined distribution on the coordinates defining the latent variable Z. Fig. 4 is a diagram for explaining the predetermined distribution. For ease of understanding, Fig. 4 shows the distribution on the X axis. The vertical axis of Fig. 4 is the number of data N.
[0025] As shown in Fig. 4, the generative model 34 is trained using a plurality of existing data DTa corresponding to each class so as to obtain a predetermined two-dimensional normal distribution for each class. In two-dimensional coordinates, the normal distribution shown in Fig. 4 is obtained on any line passing through the center coordinate Ce(0,0) of the center point PCe.
[0026] As shown in FIG. 3, the first conversion unit 21 executes a first conversion process for converting multiple pieces of existing data DTa corresponding to classes into existing coordinate data DAa for the trained generative model 34 by inputting the data to the encoder unit 343. In this embodiment, the first conversion unit 21 converts the multiple pieces of existing data DTa into a predetermined distribution on the coordinate system, thereby generating existing coordinate data DAa. Each piece of existing coordinate data DAa is a latent variable Z, which is a coordinate on the X-axis and Y-axis. Meanwhile, the second conversion unit 22 executes a second conversion process for converting the extended coordinate data DAb, which is represented by the latent variable Z generated by the data generation unit 24, into the same data format as the existing data DTa, thereby generating extended data DTb.
[0027] As shown in FIG. 3, coordinates Ce(0,0) are the center point PCe. The center point PCe is the center point of a predetermined distribution. This center point PCe is the point in the predetermined distribution that is the mode of the number of data items among the plurality of existing coordinate data DAa and the average value of the plurality of existing coordinate data DAa. Note that if the type of the predetermined distribution is a distribution other than a normal distribution, the center point PCe may be, for example, the mode of the number of data items or a reference point that takes on another value.
[0028] Fig. 5 is a flowchart showing the data generation process executed by the data generating device 10. Fig. 6 is a first diagram for explaining the data generation process. Fig. 7 is a second diagram for explaining the data generation process.
[0029] 5, in step S10, the first conversion unit 21 performs a first conversion process to convert the plurality of existing data DTa into a predetermined coordinate dimension and generate a plurality of existing coordinate data DAa. Step S10 is executed for each class of the plurality of existing data DTa using the corresponding generative model 34.
[0030] Next, in step S20, the determination unit 23 executes a determination process to determine a generation region Rga for generating the extended coordinate data DAb. The determination unit 23 executes the determination process for each class. As shown in FIG. 3 , the determination unit 23 identifies the position of a specific existing coordinate data DAap among the multiple existing coordinate data DAa that is farthest from a predetermined distribution center point PCe, and determines the generation region Rga so that all of the multiple existing coordinate data DAa are included. In this embodiment, the determination unit 23 determines, as the generation region Rga, a circular region surrounded by a circle having a center PCe, with the radius R being the distance between the predetermined distribution center point PCe of the multiple existing coordinate data DAa and the position of the specific existing coordinate data DAap. This allows the determination unit 23 to easily determine the generation region Rga. Note that the generation region Rga is not limited to a circular region surrounded by a circle, and may be a region surrounded by another shape. For example, the other shape may be a polygonal shape such as a rectangle. For example, the determination unit 23 may determine, on the coordinate system, a convex polygonal region having the smallest area that includes all of the plurality of existing coordinate data DAa, as the generation region Rga.
[0031] In another embodiment, the determination unit 23 may determine, as the generation region Rgat, a circular region surrounded by a circle centered on the center point PCe, with the radius being the distance R between a predetermined distribution center point PCe and the position of the specific existing coordinate data DAap plus an expansion value α determined based on the variance values DV of the multiple existing coordinate data DAa. The variance value DV is, for example, the average value of a first variance value on the X axis and a second variance value on the Y axis. Note that the expansion value α does not have to be determined based on the variance value DV and may be a predetermined value. By taking the expansion value α into consideration, the extended coordinate data DAb can be generated taking into account an area outside the area where the existing coordinate data DAa is located.
[0032] 5, in step S30 after step S20, the data generation unit 24 executes a generation process to generate extended coordinate data DAb so that the bias in data distribution within the generation area Rga is smaller than before the generation of the extended coordinate data DAb. In this embodiment, the bias in data distribution is represented by a data density ρ on the coordinates. In step S30, the data generation unit 24 divides the generation area Rga into a plurality of partitioned areas, and executes a generation process so that the difference in data density ρ of the existing coordinate data DAa among the respective partitioned areas is smaller than before the generation of the extended coordinate data.
[0033] As shown in FIG. 6, in step S20, the data generation unit 24 divides the generation region Rga into multiple defined regions Rga1, Rga2, and Rga3. In this embodiment, the multiple defined regions Rga1, Rga2, and Rga3 are defined according to their distance from the center point PCe. The first defined region Rga1 is a region within a distance of radius r from the center point PCe. The second defined region Rga2 is a region whose distance from the center point PCe is greater than r and is equal to or less than 2r, which is twice the radius r. The third defined region Rga3 is a region whose distance from the center point PCe is greater than 2r and is equal to or less than 3r, which is three times the radius r. The value 3r is the same as the radius R shown in FIG. 3.
[0034] As shown in FIG. 7, in step S30, the data generation unit 24 performs a generation process so that the data density ρ of each of the multiple partitioned regions Rga1 to Rga3 is the same. In the example shown in FIG. 7, extended coordinate data DAb is generated in each of the multiple partitioned regions Rga1 to Rga3 so that the data density ρ of each of the multiple partitioned regions Rga1 to Rga3 is equal to the data density ρt. The data density ρt is a value obtained by dividing the number of pieces of existing coordinate data DAa and the number of pieces of generated extended coordinate data DAb for each of the multiple partitioned regions Rga1 to Rga3 by the area of the multiple partitioned regions Rga1 to Rga3. In FIG. 7, the extended coordinate data DAb is generated as shown by single hatching. In step S20, the data generation unit 24 generates multiple candidate extended coordinate data, which are latent variables Z, using uniform random numbers. The data generation unit 24 then generates the extended coordinate data DAb by extracting, from the multiple candidate extended coordinate data, the extended coordinate data DAb for each of the multiple partitioned regions Rga1 to Rga3 by an amount equal to an increase in the number of pieces of existing coordinate data DAa required to achieve the data density ρt. The area on the coordinate system for generating multiple candidate extended coordinate data using uniform random numbers may be the entire area or may be within the generation area Rga. In this way, the data generating unit 24 can easily generate extended coordinate data DAb by using uniform random numbers.
[0035] The data density ρt is set so that the difference in the number of data items for each class is equal to or less than a predetermined class threshold. The number of data items for each class is the total number of the existing coordinate data items DAa and the extended coordinate data items DAb for each class. In this embodiment, the class threshold is set to "zero." That is, the generation process executed by the data generation unit 24 generates the extended coordinate data items DAb so that the number of data items for each class is the same. The generated extended coordinate data items DAb may be stored in the storage device 30 in association with the classes.
[0036] 5, in step S40, the second conversion unit 22 executes the second conversion process. That is, the second conversion unit 22 generates multiple pieces of extended coordinate data DAb by inputting the multiple pieces of extended coordinate data DAb to the decoder unit 346 of the generative model 34 corresponding to each class. The multiple pieces of generated extended data DTb are associated with the classes and stored in the storage device 30 as an extended data group DTbG. The existing data group DTaG and the extended data group DTbG stored in the storage device 30 are used for training the machine learning model as described above.
[0037] According to the first embodiment, the data generation unit 24 generates the extended coordinate data DAb so as to reduce the bias in data distribution within the generation region Rga, thereby reducing the possibility of bias in the distribution of data contained in the data group DTG. As a result, for example, when a machine learning model is trained using the data group DTG, a trained model 33 with improved prediction accuracy can be generated. Furthermore, according to this embodiment, since the possibility of bias in data distribution can be reduced, there is no need to generate more data than necessary and extract data suitable for training from that data. Furthermore, according to the first embodiment, the extended coordinate data DAb is generated so that the difference in the number of data items per class is equal to or less than the class threshold. As a result, for example, when a machine learning model is trained using data from each class, prediction accuracy can be improved.
[0038] 7, the data generation unit 24 executes the generation process so that the difference in data density ρ between the multiple partitioned regions Rga1, Rga2, and Rga3 is smaller than before the generation of the extended coordinate data DAb. By reducing the difference in data density ρ, the possibility of bias occurring in the distribution of data in the data group DTG can be reduced. In particular, in the first embodiment, the generation process is executed so that the difference in data density ρ between the multiple partitioned regions Rga1, Rga2, and Rga3 is the same, further reducing the possibility of bias occurring in the distribution of data in the data group DTG.
[0039] According to the first embodiment, the determination unit 23 determines the generation area Rga so that all of the multiple existing coordinate data DAa are included, as shown in Fig. 3. This allows the extended coordinate data DAb to be generated using all of the multiple existing coordinate data DAa, thereby making it possible to effectively use the existing coordinate data DAa.
[0040] B. Second embodiment: 8 is a diagram for explaining the data generation process of the second embodiment. The second embodiment differs from the first embodiment in the content of the determination process for determining the generation area and the content of the generation process for generating extended coordinate data DAb. The other configurations and processes are the same between the first and second embodiments, so descriptions of similar configurations will be omitted as appropriate.
[0041] In the second embodiment, the determination unit 23 determines, for each class, a generation region Rgap on the coordinate system for generating extended coordinate data DAb as follows: That is, the determination unit 23 determines, as the generation region Rgap, a region surrounded by a specific shape that includes a predetermined number Np or more of multiple existing coordinate data DAa on the coordinate system. The predetermined number Np is, for example, a number such that the ratio of the number of existing coordinate data DAa in the generation region Rgap to the total number of multiple existing coordinate data DAa for each class is 80% or more, preferably 90% or more. In this embodiment, as shown in FIG. 8 , the determination unit 23 determines, as the generation region Rgap, a region surrounded by a specific shape that includes all of the multiple existing coordinate data DAa on the coordinate system and is surrounded by a minimum specific shape that forms the smallest area among the specific shapes. In this embodiment, the specific shape and the minimum specific shape are convex polygonal shapes. Note that, in other embodiments, the specific shape and the minimum specific shape may be other shapes, such as a circle. In FIG. 8, the region surrounded by a rectangle connecting the four points of the existing coordinate data DAa1, DAa2, DAa3, and DAam is the generation region Rgap. By defining the region surrounded by a specific shape as the generation region Rgap, the generation region Rgap can be easily determined. In particular, by defining the region surrounded by a minimum specific shape, which is the shape that forms the smallest area among the specific shapes, as the generation region Rgap, it is possible to prevent the generation region Rgap from becoming unnecessarily large. This reduces the possibility of generating more extended coordinate data DAb than necessary.
[0042] In the second embodiment, in the generation process of step S30, the data generation unit 24 generates the extended coordinate data DAb so that the bias in data distribution within the generation region Rgap is smaller than before the generation of the extended coordinate data DAb. In the second embodiment, the bias in data distribution is represented by the variance of the inter-data distance Ds, which is the distance between two adjacent data on the coordinate system. The two adjacent data are one data and another data that is closest to the one data. The data generation unit 24 generates the extended coordinate data DAb so that the variance of the multiple inter-data distances Ds within the generation region Rgap is smaller than before the generation of the extended coordinate data DAb. In this embodiment, the data generation unit 24 generates the extended coordinate data DAb so that multiple, i.e., all, inter-data distances Ds are the same within the generation region Rgap. Furthermore, the data generation unit 24 determines the number of extended coordinate data DAb to be generated so that the total number of existing coordinate data DAa and extended coordinate data DAb is the same for each class. In the second embodiment, as in the first embodiment, the generation area Rgap may be divided into multiple partitioned areas, and the generation process may be performed so that the difference in data density between the multiple partitioned areas is smaller than before the generation of the extended coordinate data DAb.
[0043] In the second embodiment, after the generation process is executed, the second conversion process of step S40 is executed as in the first embodiment. Also, as in the first embodiment, in the second embodiment, learning of a machine learning model is executed using the data group DTG.
[0044] The second embodiment achieves the same effects as the first embodiment in that it has the same configuration and processing. The data generation unit 24 generates extended coordinate data DAb so as to reduce the bias in data distribution within the generation region Rgap, thereby reducing the possibility of bias in the distribution of data held by the data group DTG. In particular, the second embodiment generates extended coordinate data DAb so as to reduce the variance of the distances Ds between multiple pieces of data, thereby reducing the possibility of bias in the distribution of data held by the data group DTG. Furthermore, the second embodiment performs the generation process so that the distances Ds between multiple pieces of data are the same, thereby further reducing the possibility of bias in the distribution of data held by the data group DTG.
[0045] C. Other Embodiments: C-1. Alternative embodiment 1: In each of the above embodiments, as shown in FIG. 1, the data generating device 10 has functional units related to machine learning models, such as the trained model 33 and the model generating unit 26, but it does not have to have these functional units.
[0046] C-2. Alternative embodiment 2: In each of the above embodiments, the data in the storage device 30 may be stored in another device or on the cloud instead of the data generating device 10.
[0047] D. Other forms: The present disclosure is not limited to the above-described embodiments and can be realized in various forms without departing from the spirit thereof. For example, the present disclosure can also be realized in the following aspects. The technical features in the above embodiments corresponding to the technical features in each aspect described below can be appropriately replaced or combined to solve some or all of the problems of the present disclosure or to achieve some or all of the effects of the present disclosure. Furthermore, if a technical feature is not described as essential in this specification, it can be appropriately deleted.
[0048] (1) According to a first aspect of the present disclosure, there is provided a data generation device. The data generation device includes: a determination unit that executes a determination process to determine a generation area for generating extended coordinate data on the coordinate system based on positions on the coordinate system in a plurality of existing coordinate data obtained by converting each of a plurality of existing data into a predetermined coordinate dimension; and a data generation unit that executes a generation process to generate the extended coordinate data so that the bias in data distribution within the generation area is smaller than before the extended coordinate data was generated. According to this aspect, the data generation unit generates the extended coordinate data so that the bias in data distribution within the generation area is smaller, thereby reducing the possibility of bias in data distribution.
[0049] (2) In the above embodiment, the bias in the data distribution may be represented by a data density on the coordinates, and the data generator may divide the generation area into a plurality of partitioned areas and execute the generation process so that a difference in the data densities of the plurality of partitioned areas is smaller than that before the generation of the extended coordinate data. According to this embodiment, by reducing the difference in data density, it is possible to reduce the possibility of bias in the data distribution.
[0050] (3) In the above aspect, the data generation unit may perform the generation process so that the data density of each of the plurality of divided areas is the same. According to this aspect, it is possible to further reduce the possibility of bias in data distribution.
[0051] (4) In the above aspect, the system may further include a first conversion unit that executes a first conversion process to convert each of the plurality of existing data into the existing coordinate data expressed in the coordinate dimension. According to this aspect, the first conversion unit can convert the existing data into the existing coordinate data expressed in the coordinate dimension.
[0052] (5) In the above aspect, the first conversion unit may convert the plurality of existing data to have a predetermined distribution on the coordinate system, thereby generating the plurality of existing coordinate data. According to this aspect, the plurality of existing coordinate data having a predetermined distribution can be generated.
[0053] (6) In the above embodiment, the determiner may identify a position of specific existing coordinate data among the plurality of existing coordinate data that is farthest from the predetermined center point of distribution, and determine the generation area so that all of the plurality of existing coordinate data are included. According to this embodiment, the extended coordinate data can be generated using all of the plurality of existing coordinate data, so that the existing coordinate data can be used effectively.
[0054] (7) In the above aspect, the coordinates are two-dimensional coordinates, and the determiner may determine, as the generation area, a circular area surrounded by a circle having a radius equal to the distance between the center point of the distribution and the position of the specific existing coordinate data, with the center point as the center. According to this aspect, the generation area can be easily determined.
[0055] (8) In the above aspect, the coordinates are two-dimensional coordinates, and the determiner may determine, as the generation area, a circular area surrounded by a circle having a center at the center point, the circle having a radius obtained by adding an extension value determined based on variance values of the plurality of existing coordinate data to the distance between the center point of the distribution and the position of the specific existing coordinate data. According to this aspect, it is possible to generate extended coordinate data taking into account areas outside the area where the existing coordinate data are located.
[0056] (9) In the above aspect, the data generation unit may generate a plurality of candidate extended coordinate data using uniform random numbers, and extract the extended coordinate data from the plurality of candidate extended coordinate data, thereby generating the extended coordinate data. According to this aspect, the extended coordinate data can be easily generated using uniform random numbers.
[0057] (10) In the above aspect, the bias in the data distribution may be represented by a variance value of a data distance, which is the distance between two adjacent data on the coordinate, and the data generator may execute the generation process so that the variance value of the distances between the plurality of data within the generation area is smaller than that before the generation of the extended coordinate data. According to this aspect, by generating the extended coordinate data by the data generator so that the variance value of the distances between the plurality of data is smaller, it is possible to reduce the possibility of bias in the data distribution.
[0058] (11) In the above aspect, the data generation unit may execute the generation process so that the distances between the plurality of pieces of data are the same. According to this aspect, it is possible to further reduce the possibility of bias occurring in the data distribution.
[0059] (12) In the above aspect, the determination unit may determine, as the generation area, an area on the coordinate system that is surrounded by a specific shape and that includes a predetermined number or more of the plurality of existing coordinate data. According to this aspect, the generation area can be easily determined.
[0060] (13) In the above aspect, the determiner may determine, as the generation area, an area surrounded by the specific shape that includes all of the plurality of existing coordinate data on the coordinate system and that is surrounded by a minimum specific shape that is a shape that forms a smallest area among the specific shapes. According to this aspect, it is possible to prevent the generation area from becoming larger than necessary.
[0061] (14) In the above embodiment, the first conversion unit may perform the first conversion process for each of the multiple existing data classified into multiple classes, the determination unit may perform the determination process for each of the classes, and the generation process performed by the data generation unit may further include generating the extended coordinate data so that a difference in the number of data items for each of the classes is equal to or less than a predetermined class threshold. According to this embodiment, the difference in the number of data items for each of the classes can be equal to or less than the class threshold. This can improve prediction accuracy, for example, when training a machine learning model using data from each class.
[0062] (15) In the above aspect, the device may further include a second conversion unit that executes a second conversion process to convert the extended coordinate data generated by the data generation unit into the format of the existing data to generate extended data. According to this aspect, the extended data can be easily generated from the extended coordinate data using the second conversion unit.
[0063] (16) In the above aspect, the present invention may further include a model generation unit that trains a machine learning model using the plurality of existing data and the extended data to generate a trained model, and a model use unit that inputs target data to the trained model and outputs a result. According to this aspect, it is possible to output a result for the target data using the trained model with improved prediction accuracy.
[0064] (17) According to a second aspect of the present disclosure, there is provided a computer program. The computer program causes a computer to implement a determination function that executes a determination process to determine a generation area in which extended coordinate data is to be generated on the coordinate system based on the coordinate positions of a plurality of existing coordinate data obtained by converting each of a plurality of existing data into a predetermined coordinate dimension, and a data generation function that executes a generation process to generate the extended coordinate data so that the bias in data distribution within the generation area is smaller than before the extended coordinate data was generated. According to this aspect, the data generation function generates the extended coordinate data so that the bias in data distribution within the generation area is smaller, thereby reducing the possibility of bias in data distribution.
[0065] The present disclosure may be realized in various forms other than those described above, such as a data generation method or a non-transitory storage medium having a computer program recorded thereon. [Explanation of symbols]
[0066] 10...data generation device, 20...processor, 21...first conversion unit, 22...second conversion unit, 23...determination unit, 24...data generation unit, 26...model generation unit, 28...model use unit, 29...output unit, 30...storage device, 32...program, 33...trained model, 34...generative model, 34G...generative model group, 40...interface, 50...output device, 52...input device, 343...encoder unit, 346...decoder unit, Ce...center coordinates, DAa, DAa1, DAa2, DAa3, DAam...existing coordinate data, DAb...extended coordinate data, DTG...data group, DTa...existing data, DTaG...existing data group, DTb...extended data, DTbG...extended data group, DTt...data, Ds...distance between data, N...number of data, PCe...center point, R...radius, Rga, Rgap, Rgat...generation area, Rga1...first partition area, Rga2...second partition area, Rga3...third partition area
Claims
1. A data generating device, a determination unit that executes a determination process to determine a generation area in which extended coordinate data is to be generated on the coordinates, based on positions on the coordinates in a plurality of existing coordinate data obtained by converting each of a plurality of existing data into a predetermined coordinate dimension; a data generating unit that executes a generating process to generate the extended coordinate data so that a bias in data distribution within the generation area is smaller than that before the extended coordinate data is generated.
2. 2. The data generating device according to claim 1, the bias in the data distribution is represented by a data density on the coordinates; The data generation unit divides the generation area into a plurality of partitioned areas, and performs the generation process so that a difference in the data density between the plurality of partitioned areas is smaller than before the extended coordinate data was generated.
3. 3. The data generating device according to claim 2, The data generation unit executes the generation process so that the data density of each of the plurality of partitioned areas is the same.
4. 2. The data generating device according to claim 1, further comprising: a first conversion unit that executes a first conversion process to convert each of the plurality of existing data into the existing coordinate data expressed in the coordinate dimension;
5. 5. The data generating device according to claim 4, The first conversion unit converts the plurality of existing data to have a predetermined distribution on the coordinate system, thereby generating the plurality of existing coordinate data.
6. 6. The data generating device according to claim 5, The determination unit identifies the position of a specific existing coordinate data item among the plurality of existing coordinate data items that is farthest from the predetermined center point of distribution, and determines the generation area so that all of the plurality of existing coordinate data items are included.
7. 7. The data generating device according to claim 6, the coordinates are two-dimensional coordinates, The determination unit determines, as the generation area, a circular area surrounded by a circle having a radius equal to the distance between the center point of the distribution and the position of the specific existing coordinate data and a center point of the distribution.
8. 7. The data generating device according to claim 6, the coordinates are two-dimensional coordinates, The determination unit determines, as the generation area, a circular area surrounded by a circle centered on the center point, with a radius being the distance between the center point of the distribution and the position of the specific existing coordinate data plus an expansion value determined based on the variance values of the multiple existing coordinate data.
9. 3. The data generating device according to claim 2, The data generation unit generates a plurality of candidate extended coordinate data using uniform random numbers, and extracts the extended coordinate data from the plurality of candidate extended coordinate data, thereby generating the extended coordinate data.
10. 2. The data generating device according to claim 1, the deviation of the data distribution is represented by a variance value of a data distance, which is a distance between two adjacent data on the coordinate system; The data generation unit executes the generation process so that, within the generation region, the variance value of the distances between the plurality of pieces of data becomes smaller than that before the extended coordinate data was generated.
11. 11. The data generating device according to claim 10, The data generation unit executes the generation process so that the distances between the plurality of data are the same.
12. 2. The data generating device according to claim 1, The determination unit determines, as the generation area, an area surrounded by a specific shape on the coordinate system that includes a predetermined number or more of the plurality of existing coordinate data.
13. 13. The data generating device according to claim 12, The determination unit determines, as the generated area, an area surrounded by the specific shape that includes all of the multiple existing coordinate data on the coordinate system and is surrounded by a minimum specific shape that is a shape that forms the smallest area among the specific shapes.
14. 5. The data generating device according to claim 4, the first conversion unit performs the first conversion process for each of the plurality of existing data classified into a plurality of classes; the determination unit executes the determination process for each of the classes; The generation process executed by the data generation unit further includes generating the extended coordinate data so that a difference in the number of data items for each class is equal to or less than a predetermined class threshold.
15. The data generating device according to any one of claims 1 to 14, further comprising: a second conversion unit that executes a second conversion process to convert the extended coordinate data generated by the data generation unit into a format of the existing data to generate extended data;
16. 16. The data generating device according to claim 15, further comprising: a model generation unit that trains a machine learning model using the plurality of existing data and the extended data to generate a trained model; A data generation device comprising: a model usage unit that inputs target data into the trained model and outputs the results.
17. A computer program comprising: a determination function that executes a determination process to determine a generation area for generating extended coordinate data on the coordinates based on positions on the coordinates of a plurality of existing coordinate data obtained by converting each of a plurality of existing data into a predetermined coordinate dimension; a data generation function that executes a generation process to generate the extended coordinate data so that the bias in data distribution within the generation area is smaller than before the extended coordinate data was generated; and a computer program that causes a computer to realize the above functions.