Training data generation program, training data generation method, and training data generation apparatus

Abstract

To generate training data that can track concept drift even when there is an imbalance in data between multiple classes.SOLUTION: A training data generation device determines a conversion parameter on the basis of the number of pieces of data, of a plurality of pieces of training data, which belongs to a first class, a second class, and a third class, respectively, and on the basis of the conversion parameter, combines first training data in the first class, second training data in the second class, and third training data in the third class to generate fourth training data.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a technique for generating training data.

Background Art

[0002] In an AI (Artificial Intelligence) system, for example, due to reasons such as differences between teacher data and operation data during development, and changes in statistical trends of operation data, the prediction accuracy of a machine learning model may decrease and become obsolete during operation. Therefore, the prediction accuracy of the machine learning model is regularly monitored, and if the accuracy has decreased, the machine learning model is retrained. The change in the quality of data input to the AI system over time can be referred to as concept drift.

[0003] For example, a method is known in which pseudo labels are generated by clustering input data in a feature space, and the machine learning model is retrained based on the pseudo labels to follow concept drift.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Patent Document 3

Patent Document 4

Summary of the Invention

Problems to be Solved by the Invention

[0005] However, conventional methods that generate pseudo-labels by clustering such input data in the feature space and then retrain a machine learning model based on these pseudo-labels to track concept drift struggle to track data imbalances between multiple classes.

[0006] In one aspect, the present invention aims to generate training data that can track concept drift even when there is a data imbalance between multiple classes. [Means for solving the problem]

[0007] Therefore, this training data generation program causes the computer to perform the following process: determine conversion parameters based on the number of data points belonging to each of the first, second, and third classes among the multiple training data sets; and generate fourth training data by combining the first training data of the first class, the second training data of the second class, and the third training data of the third class based on the conversion parameters. The process for determining the transformation parameters includes determining the transformation parameters based on a relaxation probability density function defined to resolve the imbalance in the ratio of the number of data belonging to the first class to the number of data belonging to the second class to the number of data belonging to the third class. The process for generating the fourth training data includes linearly combining the first training data with the second training data and the third training data, which have the shortest sum of mutual distances including the first training data, using the transformation parameters. [Effects of the Invention]

[0008] According to one embodiment, it is possible to generate training data that can track concept drift even when there is a data imbalance between multiple classes. [Brief explanation of the drawing]

[0009] [Figure 1] This diagram schematically shows the configuration of an information processing device as an example of an embodiment. [Figure 2] This figure illustrates the inference results obtained by a machine learning model on input data in an information processing device, as an example of an embodiment. [Figure 3] This figure illustrates the processing of a probability density function generation unit in an information processing device as an example of an embodiment. [Figure 4]This is a diagram illustrating the processing of a data selection unit in an information processing device as an example of an embodiment. [Figure 5] This is a diagram illustrating the processing of the data synthesis unit in an information processing device as an example of an embodiment. [Figure 6] This is a flowchart illustrating the processing in an information processing device as an example of an embodiment. [Figure 7] This figure illustrates the hardware configuration of an information processing device as an example of an embodiment. [Modes for carrying out the invention]

[0010] The following describes embodiments of the training data generation program, training data generation method, and training data generation apparatus with reference to the drawings. However, the embodiments shown below are merely illustrative, and there is no intention to exclude various modifications or applications of techniques not explicitly shown in the embodiments. In other words, these embodiments can be implemented with various modifications without departing from their spirit. Furthermore, each figure does not mean that only the components shown are included, but may also include other functions, etc.

[0011] (A) Configuration Figure 1 is a schematic diagram showing the configuration of an information processing device 1 as an example of an embodiment.

[0012] As shown in Figure 1, the information processing device 1 has the function of a pseudo-data generation unit 100.

[0013] The pseudo-data generation unit 100 generates pseudo-data to resolve imbalances between multiple classes. The pseudo-data generated by the pseudo-data generation unit 100 is used as training data in training a machine learning model (machine learning). In other words, the pseudo-data generation unit 100 implements the function of generating training data.

[0014] The machine learning model may perform class classification using an unsupervised machine learning clustering algorithm. Further, the machine learning model may perform multi-class classification that classifies the input data into any one of three or more classes (multi-class). The machine learning model may perform deep distance learning and density-based clustering.

[0015] A class obtained by the machine learning model making an inference on the input data is referred to as an inference class or simply a class. and Also, regarding the inference class obtained as a result of the machine learning model making an inference on the input data, for convenience, this input data may be expressed as the input data of the inference class.

[0016] As shown in FIG. 1, the pseudo data generation unit 100 has functions as an inference result acquisition unit 101, a probability density function generation unit 102, a data selection unit 103, and a data synthesis unit 104.

[0017] The inference result acquisition unit 101 acquires the inference result (class classification result) for the training data by the machine learning model.

[0018] The inference result acquisition unit 101 may input a plurality of input data (training data) to the machine learning model to perform an inference. Also, the inference result acquisition unit 101 may acquire the inference result by having the machine learning model perform an inference on the input data (training data) in advance and reading out the inference result stored in a storage device (not shown) or the like.

[0019] FIG. 2 is a diagram illustrating an inference result for input data by a machine learning model in the information processing apparatus 1 as an example of an embodiment.

[0020] In the example shown in Figure 2, we see the results of inference performed on a machine learning model that performs multi-class classification (multi-class classification) to estimate whether the input image data belongs to a dog (DOG), cat (CAT), or rabbit (RABBIT), using 160 data points. Specifically, it shows an example where the inference classes are three: dog (DOG), cat (CAT), and rabbit (RABBIT).

[0021] In the example shown in Figure 2, after inputting 160 data points, 100 operational data points were inferred to be cats (CAT), 50 operational data points were inferred to be dogs (DOG), and 10 operational data points were inferred to be rabbits (RABBIT).

[0022] The probability density function generation unit 102 generates a probability density function that mitigates the imbalance in the proportions of each class in the inference results, based on the inference results obtained by the inference result acquisition unit 101 on the operational data by the machine learning model. The input data during the operational phase in which the machine learning model is in operation can be called operational data. Note that all data used for training (including retraining) the machine learning model is also training data, so operational data is also included in the training data.

[0023] Figure 3 is a diagram illustrating the processing of the probability density function generation unit 102 in the information processing device 1 as an example of an embodiment.

[0024] Figure 3 schematically shows the imbalance state and the mitigation probability density function calculated based on the inference results for the operational data exemplified in Figure 2.

[0025] In Figure 3, symbol A represents an example of an imbalance in the proportions of each class in the inference results for operational data from a machine learning model. Symbol B schematically represents a relaxation probability density function used to alleviate the imbalance shown in symbol A.

[0026] The probability density function generator 102 defines a Dirichlet distribution that takes into account the inference result ratio of each class and generates a probability density function that mitigates the imbalance.

[0027] The probability density function generator 102 defines a probability density function that resolves the imbalance between multiple classes in the inference result.

[0028] In the example shown by code A in Figure 3, there is an imbalance where a large amount of data is estimated to be cat (CAT). To resolve this imbalance, the probability density function generator 102 defines a probability density function that estimates a large amount of dog (DOG) or rabbit (RABBIT), as shown by code B in Figure 3.

[0029] The probability density function generator 102 uses the Dirichlet distribution, a multivariate extension of the beta distribution, to set up a function that can arbitrarily obtain multiple event probabilities. The Dirichlet distribution and its probability density function are known, and a detailed explanation is omitted.

[0030] Furthermore, the probability density function generator 102 uses standardization methods such as log1p to mitigate (smooth out) extreme value differences while considering the imbalance state. The subsequent generation of the relaxation function uses the inverse value after standardization. 。

[0031] This prevents the function from becoming one that returns a very limited probability distribution when generating a Dirichlet distribution that takes into account the inference result ratio of each class, if the output value is used directly in the case of extreme imbalance.

[0032] The probability density function generation unit 102 uses the defined probability density function to generate a set of synthesis parameters for use by the data synthesis unit 104, described later, to synthesize pseudodata. The synthesis parameter values ​​are combinations of probability distribution values ​​(parameter values) corresponding to each of the multiple classes that make up the multiclass (for example, cat, dog, rabbit), and are, for example, the values ​​shown below.

[0033] Cat: 0.010760 Dog: 0.645617 Rabbit: 0.343623

[0034] A synthesis parameter set is a combination of transformation parameters used to resolve imbalances between multiple classes in the inference result, and is determined by the probability density function (probability function) described above. The sum of the transformation parameters in a single synthesis parameter set is 1. A synthesis parameter set represents a synthesis ratio set for multiple classes that make up a multiclass (e.g., cat, dog, rabbit). A synthesis parameter set can also be called a stochastic state.

[0035] The probability density function generator 102 generates multiple composite parameter sets based on the probability density function described above. The probability density function generator 102 determines the composite parameter set (transformation parameters) based on the number of data points belonging to the first class, second class, and third class from the multiple training data. The first class, second class, and third class may be, for example, class "rabbit", class "cat", and class "dog".

[0036] The probability density function generator 102 generates a number of composite parameter sets sufficient to resolve imbalances between multiple classes in the inference results.

[0037] The number sufficient to resolve the imbalance between multiple classes in the inference results may be, for example, the sum of the differences in the number of inferences for each of the remaining classes relative to the number of inferences for the class with the highest number of inference results (the most frequent class).

[0038] For example, in the example shown in Figure 2, the most frequent class, "cat," has 100 inferences. In contrast, the difference between this and the class "dog," which has 50 inferences, is 50, and the difference between this and the class "rabbit," which has 10 inferences, is 90. Therefore, the number of inferences sufficient to resolve the imbalance between multiple classes in the inference results is 140 (=50+90), and the probability density function generator 102 generates 140 types of composite parameter sets.

[0039] In other words, the probability density function generation unit 102 uses the defined probability density function to generate a set of composite parameters large enough to resolve the imbalance between multiple classes in the inference result.

[0040] The generated composite parameter set is stored, for example, in a predetermined storage area of ​​the storage device 13 (see Figure 7).

[0041] The data selection unit 103 selects data from multiple training datasets to be used by the data synthesis unit 104 (described later) to generate pseudo-data, and creates a dataset for pseudo-data generation.

[0042] The data selection unit 103 selects multiple representative data for each of the multiple inference classes. For example, the data selection unit 103 may determine the top n input data (where n is a natural number greater than or equal to 2) with the highest confidence levels from among the multiple input data for each class as the representative data for each class. Alternatively, when selecting representative data, the data selection unit 103 may set a priority order based, for example, on the confidence value of the inference result.

[0043] The method used by the data selection unit 103 to determine representative data is not limited to this and can be modified as appropriate. For example, the data selection unit 103 may determine input data with a confidence level equal to or greater than a predetermined threshold as representative data for each class.

[0044] Furthermore, the data selection unit 103 calculates the distance between each representative data point of each class and each data point of the other classes, and determines the combination of data points that has the shortest total distance between the data points, including the representative data point, as the dataset for generating pseudo-data.

[0045] In other words, when determining the dataset for generating pseudo-data, the data selection unit 103 selects the nearest neighbor data in each other class for the representative data. This makes it less likely for the data synthesis unit 104, described later, to generate pseudo-data by synthesizing multiple data included in the pseudo-data generation dataset, and enables the generation of the intended pseudo-data. Furthermore, it prevents the selection of unusual data combinations, thereby improving reliability. In addition, it is possible to reduce the discrepancy between the pseudo-data and the existing machine learning model and data by considering the entire linear combination target, not just the source.

[0046] The class to which the representative data belongs may be called the first class. The classes other than the first class may be called the second class and the third class, respectively. The data selection unit 103 selects the training data that has the shortest total distance between each other, including the representative data (first training data), from the training data belonging to the second class and the training data belonging to the third class, respectively, and determines them to be the second training data and the third training data.

[0047] The data selection unit 103 may determine the distance between data points using, for example, Euclidean distance or cosine similarity. Furthermore, it is not limited to these methods and can be modified as appropriate.

[0048] Figure 4 is a diagram illustrating the processing of the data selection unit 103 in the information processing device 1 as an example of an embodiment.

[0049] In Figure 4, multiple input data plotted in the feature space are classified into three classes: cat (CAT), dog (DOG), and rabbit (RABBIT).

[0050] The data selection unit 103 selects data from the "CAT" class and the "DOG" class, respectively, that have the shortest sum of distances between them (see codes P1 and P2), based on the representative data of the "RABBIT" class (see code P0), and determines them as a dataset for generating pseudodata.

[0051] The data selection unit 103 selects input data P1 from the input data of class "DOG" that is closest to representative data P0. The data selection unit 103 also selects input data P2 from the input data of class "CAT" that is closest to representative data P0. Furthermore, it is desirable that the distance between input data P1 and input data P2 is also the shortest possible between multiple input data of class "DOG" and multiple input data of class "CAT".

[0052] In this case, the class "RABBIT" corresponds to the first class, and the representative data indicated by the code P0 in Figure 4 corresponds to the first training data. The classes "CAT" and "DOG" correspond to the second and third classes, respectively. Furthermore, the training data indicated by the code P2 in Figure 4 corresponds to the second training data, and the training data indicated by the code P1 corresponds to the third training data.

[0053] The data synthesis unit 104 synthesizes multiple data points included in the pseudo-data generation dataset determined by the data selection unit 103 to generate pseudo-data. The data synthesis unit 104 generates pseudo-data by linearly combining multiple data points included in the pseudo-data generation dataset.

[0054] In this process, the data synthesis unit 104 applies the synthesis parameter set configured by the probability density function generation unit 102 to multiple data points included in the pseudo-data generation dataset. Specifically, when the data synthesis unit 104 synthesizes data from each class constituting the pseudo-data generation dataset to generate pseudo-data, it reflects the values ​​(parameter values) corresponding to each class in the synthesis parameter set as weights in the data from each class constituting the pseudo-data generation dataset.

[0055] The data synthesis unit 104 generates pseudo-data (fourth training data) by synthesizing the first training data of class "cat" (first class), the second training data of class "dog" (second class), and the third training data (representative data) of class "rabbit" (third class) based on a synthesis parameter set (conversion parameters).

[0056] Figure 5 is a diagram illustrating the processing of the data synthesis unit 104 in the information processing device 1 as an example of an embodiment.

[0057] Figure 5 shows an example of generating pseudo-data (see code P3) that has a high probability of being inferred to be the class "RABBIT" by combining the data indicated by codes P0, P1, and P2 as exemplified in Figure 4.

[0058] Furthermore, in the example shown in Figure 5, 50 pseudo-data sets with a high probability of being inferred as class "DOG" are generated, and 90 pseudo-data sets with a high probability of being inferred as class "RABBIT" are generated, thereby resolving the imbalance in the number of inference results among the classes "CAT", "DOG", and "RABBIT".

[0059] The data synthesis unit 104 generates pseudo-data that resolves imbalances between classes by performing data synthesis according to the selected datasets (pseudo-data generation datasets) and the obtained set of stochastic states (synthesis parameter sets).

[0060] (B) Operation The processing in the information processing device 1, as an example of the embodiment configured as described above, will be explained according to the flowchart (steps S1 to S8) shown in Figure 6.

[0061] In step S1, a machine learning model is created. In the training phase, the machine learning model is trained using the training data.

[0062] In step S2, operational data is input into the machine learning model and inference is performed. The inference result acquisition unit 101 collects the inference results. From step S2 onward, the operation phase begins.

[0063] In step S3, the probability density function generation unit 102 aggregates the number of data points for each inference class and generates a Dirichlet distribution that returns the inverse ratio of each of the multiple inference classes.

[0064] In step S4, the probability density function generator 102 obtains the necessary number of probability states (composition parameter sets) to bring the imbalanced state between multiple classes into equilibrium, based on the Dirichlet distribution.

[0065] Furthermore, in step S5, the data selection unit 103 selects multiple representative data for each of the multiple inference classes. For example, the data selection unit 103 determines the top n input data (where n is a natural number greater than or equal to 2) with the highest confidence levels as the representative data for each class.

[0066] In step S6, the data selection unit 103 calculates the distance between the representative data and each data in an inference class other than the class to which the representative data belongs.

[0067] In step S7, the data selection unit 103 selects data from each of the other inference classes for the representative data, where the sum of the distances between them is the shortest, and determines it as a dataset for generating pseudodata. The data selection unit 103 creates the necessary number of pseudodata generation datasets to bring the imbalance in class ratios among the multiple classes into equilibrium. The processes in steps S5 to S7 described above may be performed in parallel with steps S3 to S4. Furthermore, the processes in steps S5 to S7 may be performed after the completion of the processes in steps S3 to S4, or before the processes in steps S3 to S4, and can be modified as appropriate.

[0068] In step S8, the data synthesis unit 104 synthesizes multiple data points included in the pseudo-data generation dataset determined by the data selection unit 103 to generate pseudo-data. After that, the process ends.

[0069] (C) Effects As described above, according to the information processing device 1 as an example of an embodiment, the probability density function generation unit 102 aggregates the number of data points for each inference class for the inference results on operational data and generates a Dirichlet distribution that returns the inverse ratio of each of the multiple inference classes. Then, based on the Dirichlet distribution, the probability density function generation unit 102 obtains the number of probability states (composition parameter set) necessary to bring the imbalanced state between multiple classes into an equilibrium state.

[0070] Furthermore, the data selection unit 103 selects multiple representative data points for each of the multiple inference classes. The data selection unit 103 also calculates the distance between each representative data point and each data point in the inference classes other than the class to which the representative data point belongs. Then, the data selection unit 103 selects the data points from each of the other inference classes that have the shortest sum of mutual distances to the representative data point, and determines them as the dataset for pseudo-data generation.

[0071] Then, the data synthesis unit 104 synthesizes multiple data points included in the pseudo-data generation dataset to generate pseudo-data. In this process, the data synthesis unit 104 also applies the synthesis parameter set configured by the probability density function generation unit 102 to the multiple data points included in the pseudo-data generation dataset.

[0072] This makes it easy to generate pseudo-data suitable for resolving imbalances between multiple classes, and also makes it easy to generate training data that can track concept drift even when there is a data imbalance between multiple classes.

[0073] Furthermore, the probability density function generator 102 uses standardization methods such as log1p to mitigate (smooth) extreme value differences while taking imbalances into account. This prevents the function from returning a very limited probability distribution when generating a Dirichlet distribution that takes into account the inference result ratio of each class, if the output value is used directly in the case of extreme imbalances.

[0074] The data selection unit 103 selects the nearest neighbor data for each class relative to the representative data when determining the dataset for generating pseudo-data. This makes it less likely for the data synthesis unit 104 to encounter problems when synthesizing multiple data included in the pseudo-data generation dataset to generate pseudo-data, and enables the generation of the intended pseudo-data. Furthermore, it prevents the selection of unusual data combinations, thereby improving reliability. In addition, it considers the entire linear combination target, not just the source, and reduces the discrepancy between the pseudo-data and the existing machine learning model and data.

[0075] (D) Other Figure 7 is a diagram illustrating the hardware configuration of an information processing device 1 as an example of an embodiment.

[0076] The information processing device 1 includes, for example, a processor 11, memory 12, storage device 13, graphics processing device 14, input interface 15, optical drive device 16, device connection interface 17, and network interface 18 as its components. These components 11 to 18 are configured to communicate with each other via a bus 19.

[0077] The processor (control unit) 11 controls the entire information processing device 1. The processor 11 may be a multiprocessor. The processor 11 may be one of the following: CPU, MPU (Micro Processing Unit), DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), PLD (Programmable Logic Device), FPGA (Field Programmable Gate Array), or GPU (Graphics Processing Unit). Alternatively, the processor 11 may be a combination of two or more elements from among CPU, MPU, DSP, ASIC, PLD, FPGA, and GPU.

[0078] Then, the processor 11 executes a control program for the information processing device 1 (training data generation program: not shown), thereby realizing the function of a pseudo-data generation unit 100 as illustrated in Figure 1.

[0079] Furthermore, the information processing device 1 realizes its function as a pseudo-data generation unit 100 by executing a program (training data generation program, OS program) recorded on, for example, a computer-readable non-temporary recording medium. OS is an abbreviation for Operating System.

[0080] The program describing the processing to be executed by the information processing device 1 can be stored on various recording media. For example, the program to be executed by the information processing device 1 can be stored in the storage device 13. The processor 11 loads at least a portion of the program in the storage device 13 into memory 12 and executes the loaded program.

[0081] Furthermore, programs to be executed by the information processing device 1 (processor 11) can also be recorded on non-temporary portable recording media such as optical disc 16a, memory device 17a, or memory card 17c. Programs stored on portable recording media can be installed on storage device 13, for example, under control from processor 11, and then become executable. Alternatively, processor 11 can directly read and execute programs from portable recording media.

[0082] Memory 12 is a storage memory that includes ROM (Read Only Memory) and RAM (Random Access Memory). The RAM of memory 12 is used as the main memory of the information processing device 1. At least a portion of the program to be executed by the processor 11 is temporarily stored in the RAM. In addition, various data necessary for processing by the processor 11 is stored in memory 12.

[0083] The storage device 13 is a storage device such as a hard disk drive (HDD), solid state drive (SSD), or storage class memory (SCM), and stores various types of data. The storage device 13 is used as an auxiliary storage device for the information processing device 1. The storage device 13 stores the OS program, control program, and various types of data. The control program includes a training data generation program.

[0084] Furthermore, semiconductor memory devices such as SCMs and flash memory can be used as auxiliary storage devices. Additionally, RAID (Redundant Arrays of Inexpensive Disks) may be configured using multiple storage devices 13.

[0085] Furthermore, the storage device 13 may store various types of data generated when the pseudo-data generation unit 100 performs each process as described above.

[0086] A monitor 14a is connected to the graphics processing unit 14. The graphics processing unit 14 displays an image on the screen of the monitor 14a according to instructions from the processor 11. The monitor 14a can be a display device using a CRT (Cathode Ray Tube) or a liquid crystal display device.

[0087] A keyboard 15a and a mouse 15b are connected to the input interface 15. The input interface 15 transmits signals from the keyboard 15a and mouse 15b to the processor 11. Note that the mouse 15b is just one example of a pointing device, and other pointing devices can also be used. Other pointing devices include touch panels, tablets, touchpads, and trackballs.

[0088] The optical drive device 16 reads data recorded on the optical disc 16a using laser light or the like. The optical disc 16a is a portable, non-temporary recording medium on which data is recorded in a readable format by light reflection. Examples of optical discs 16a include DVD (Digital Versatile Disc), DVD-RAM, CD-ROM (Compact Disc Read Only Memory), CD-R (Recordable) / RW (ReWritable), etc.

[0089] The device connection interface 17 is a communication interface for connecting peripheral devices to the information processing device 1. For example, a memory device 17a and a memory reader / writer 17b can be connected to the device connection interface 17. The memory device 17a is a non-temporary recording medium equipped with a communication function with the device connection interface 17, such as a USB (Universal Serial Bus) memory. The memory reader / writer 17b writes data to or reads data from the memory card 17c. The memory card 17c is a card-type non-temporary recording medium.

[0090] The network interface 18 is connected to a network. The network interface 18 transmits and receives data via the network. Other information processing devices and communication devices may also be connected to the network.

[0091] Furthermore, the disclosed technology is not limited to the embodiments described above, and can be implemented in various modified forms without departing from the spirit of this embodiment. Each configuration and process of this embodiment can be selected or combined as needed.

[0092] For example, the data selection unit 103 may calculate the distance between each representative data point of each class and representative data points of other classes, and determine the combination of representative data points that results in the shortest total distance. This allows for selection in a short amount of time.

[0093] Furthermore, the above disclosure makes it possible for those skilled in the art to implement and manufacture this embodiment.

[0094] (E) Note The following additional information is disclosed regarding the embodiments described above.

[0095] (Note 1) The transformation parameters are determined based on the number of data points belonging to the first, second, and third classes in the multiple training datasets. Based on the conversion parameters, a fourth training data is generated by combining the first training data of the first class, the second training data of the second class, and the third training data of the third class. A training data generation program characterized by having a computer perform the processing.

[0096] (Note 2) The process for determining the aforementioned conversion parameters is: A relaxation probability density function is set to alleviate the imbalance between the number of data points belonging to the first class, the number of data points belonging to the second class, and the number of data points belonging to the third class, and the transformation parameters are determined based on the relaxation probability density function. A training data generation program as described in Appendix 1, characterized by including processing.

[0097] (Note 3) The first training data is selected from the training data belonging to the first class, The training data that has the shortest sum of the distances between each other, including the determined first training data, is selected from the training data belonging to the second class and the training data belonging to the third class, respectively, and determined to be the second training data and the third training data. A training data generation program according to Appendix 1 or 2, characterized in that it causes the computer to perform the processing.

[0098] (Note 4) The transformation parameters are determined based on the number of data points belonging to the first, second, and third classes in the multiple training datasets. Based on the conversion parameters, a fourth training data is generated by combining the first training data of the first class, the second training data of the second class, and the third training data of the third class. A method for generating training data, characterized in that the processing is performed by a computer.

[0099] (Note 5) The process for determining the aforementioned conversion parameters is: A relaxation probability density function is set to alleviate the imbalance between the number of data points belonging to the first class, the number of data points belonging to the second class, and the number of data points belonging to the third class, and the transformation parameters are determined based on the relaxation probability density function. The training data generation method described in Appendix 4, characterized by including processing.

[0100] (Note 6) The first training data is selected from the training data belonging to the first class, The training data that has the shortest sum of the distances between each other, including the determined first training data, is selected from the training data belonging to the second class and the training data belonging to the third class, respectively, and determined to be the second training data and the third training data. The training data generation method according to Appendix 4 or 5, characterized in that the processing is performed by the computer.

[0101] (Note 7) The transformation parameters are determined based on the number of data points belonging to the first, second, and third classes in the multiple training datasets. Based on the conversion parameters, a fourth training data is generated by combining the first training data of the first class, the second training data of the second class, and the third training data of the third class. A training data generation device characterized by having a control unit that performs processing.

[0102] (Note 8) The process for determining the aforementioned conversion parameters is: A relaxation probability density function is set to alleviate the imbalance between the number of data points belonging to the first class, the number of data points belonging to the second class, and the number of data points belonging to the third class, and the transformation parameters are determined based on the relaxation probability density function. A training data generation device according to Appendix 7, characterized by including processing.

[0103] (Note 9) The control unit, The first training data is selected from the training data belonging to the first class, The training data that has the shortest sum of the distances between each other, including the determined first training data, is selected from the training data belonging to the second class and the training data belonging to the third class, respectively, and determined to be the second training data and the third training data. A training data generation device according to Appendix 7 or 8, characterized by performing processing. [Explanation of symbols]

[0104] 1. Information Processing System 10 Computers 11. Processor (control unit) 12 memory 13 Storage device 14 Graphics Processing Unit 14a Monitor 15 Input Interfaces 15a Keyboard 15b Mouse 16 Optical drive device 16a Optical disc 17. Device connection interface 17a Memory device 17b Memory Reader / Writer 17c memory card 18 Network Interfaces 18a Network 19 bus 100 Pseudo-data generation unit 101 Probability Density Function Generator 103 Data Selection Section 104 Data Synthesis Unit

Claims

1. The transformation parameters are determined based on the number of data points belonging to the first, second, and third classes among the multiple training datasets. Based on the conversion parameters, a fourth training data is generated by combining the first training data of the first class, the second training data of the second class, and the third training data of the third class. Let the computer perform the process, The process for determining the transformation parameters includes determining the transformation parameters based on a relaxation probability density function defined to resolve the imbalance in the ratio of the number of data belonging to the first class, the number of data belonging to the second class, and the number of data belonging to the third class. The process for generating the fourth training data includes a process of linearly combining the first training data and the second and third training data, which have the shortest sum of the distances between each other, including the first training data, using the transformation parameter. A training data generation program characterized by the following features.

2. The process for determining the aforementioned conversion parameters is: The value difference in the imbalance between the number of data belonging to the first class, the number of data belonging to the second class, and the number of data belonging to the third class is mitigated using a standardization method. The relaxation probability density function is set based on the inverse value of the relaxed value. A training data generation program according to claim 1, characterized by including processing.

3. From the training data belonging to the first class, training data in which the confidence or reliability of the inference result is equal to or greater than a predetermined threshold, or the top n (n is a natural number of 2 or more) training data with high confidence or reliability, are selected as the first training data. The training data generation program according to claim 1 or 2, characterized in that it causes the computer to perform the processing.

4. The process for determining the conversion parameters includes generating a number of combinations of the conversion parameters corresponding to the first class, the second class, and the third class, based on the sum of the differences in the number of data points of the remaining classes compared to the number of data points of the class with the largest number of data points among the first class, the second class, and the third class. A training data generation program according to claim 1 or 2, characterized in that...

5. The transformation parameters are determined based on the number of data points belonging to the first, second, and third classes among the multiple training datasets. Based on the conversion parameters, a fourth training data is generated by combining the first training data of the first class, the second training data of the second class, and the third training data of the third class. The computer performs the process, The process for determining the transformation parameters includes determining the transformation parameters based on a relaxation probability density function defined to resolve the imbalance in the ratio of the number of data belonging to the first class, the number of data belonging to the second class, and the number of data belonging to the third class. The process for generating the fourth training data includes a process of linearly combining the first training data and the second and third training data, which have the shortest sum of the distances between each other, including the first training data, using the transformation parameter. A method for generating training data characterized by the following features.

6. The transformation parameters are determined based on the number of data points belonging to the first, second, and third classes among the multiple training datasets. Based on the conversion parameters, a fourth training data is generated by combining the first training data of the first class, the second training data of the second class, and the third training data of the third class. It has a control unit that performs processing, The process for determining the transformation parameters includes determining the transformation parameters based on a relaxation probability density function defined to resolve the imbalance in the ratio of the number of data belonging to the first class, the number of data belonging to the second class, and the number of data belonging to the third class. The process for generating the fourth training data includes a process of linearly combining the first training data and the second and third training data, which have the shortest sum of the distances between each other, including the first training data, using the transformation parameter. A training data generation device characterized by the following features.

Images