A method and apparatus for classifying data.

Abstract

The present invention provides a method and apparatus for classifying data. [Solution] A method for classifying input data includes the steps of: training a classification model to generate a first output value from a first mathematical formula corresponding to a classification model that classifies input data into at least one class, by disentanglement of a second mathematical formula in which a component corresponding to the label distribution of source data is disentangled; generating a second output value by reflecting information representing the label distribution of target data in the first output value; and classifying the target data into the at least one class using the second output value.

Description

[Technical Field]

[0001] This relates to a method and apparatus for classifying data. [Background technology]

[0002] In recent years, techniques for classifying input data into predetermined classes in combination with deep learning have been developing. A classification model (or classifier) ​​determines which class the input data belongs to, and even if the input data does not belong to any of the predefined classes, it classifies it into the most similar class among the predefined classes. Therefore, the accuracy of the classification model is considered a crucial factor in ensuring the integrity of services.

[0003] On the other hand, when generating classification models using deep learning technology, the accuracy of the classification model depends on the distribution of the training data. Therefore, there is a growing demand for technologies that can generate accurate classification models regardless of the distribution of the training data. [Overview of the Initiative] [Problems that the invention aims to solve]

[0004] This invention provides a method and apparatus for classifying data. It also provides a computer-readable recording medium containing a program for executing the above-described method on a computer. The technical problems to be solved are not limited to those described above, and other technical problems may exist. [Means for solving the problem]

[0005] A method for classifying data according to one embodiment includes the steps of: training a classification model to generate a first output value from a first mathematical formula corresponding to a classification model that classifies input data into at least one class, by disentanglement of a component corresponding to the label distribution of source data into a second mathematical formula; generating a second output value by reflecting information representing the label distribution of target data in the first output value; and classifying the target data into the at least one class using the second output value.

[0006] Computer-readable recording media in other embodiments include recording media on which a program for causing a computer to perform the method described above is recorded.

[0007] Another apparatus for classifying data according to other embodiments includes a memory for storing at least one program and a processor that operates by executing the at least one program, wherein the processor learns the classification model to generate a first output value by a second formula obtained by untangling a first formula corresponding to the classification model with a component corresponding to the label distribution of source data, generates a second output value by reflecting information representing the label distribution of target data in the first output value, and classifies the target data into the at least one class using the second output value. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 illustrates an example of classifying input data into at least one class. [Figure 2A] Figure 2A illustrates an example of how a classification model works during the learning and inference phases. [Figure 2B] Figure 2B illustrates an example of how a classification model works during the learning and inference phases. [Figure 3] Figure 3 illustrates an example of classification results based on the distribution of source data used for training. [Figure 4] FIG. 4 is a flowchart showing an example of a method for classifying data according to an embodiment. [Figure 5] FIG. 5 is a configuration diagram showing an example of an apparatus for classifying data according to an embodiment. [Figure 6] FIG. 6 is a diagram for explaining an example in which a second output value according to an embodiment is utilized.

BEST MODE FOR CARRYING OUT THE INVENTION

[0009] The terms used in the embodiments are, as much as possible, general terms that are currently widely used. However, this may change according to the intention or precedent of those skilled in the art in the relevant technical field, the emergence of new technologies, etc. Also, in certain cases, there are terms arbitrarily selected by the applicant. In such cases, the meaning thereof will be described in detail in the corresponding explanatory part. Therefore, the terms used in the specification should be defined based not on merely the names of the terms, but on the meaning of the terms and the content throughout the specification.

[0010] When a part of the entire specification states that a certain component "includes", this means that, unless otherwise stated to the contrary, it does not exclude other components, but may further include other components. Also, terms such as "… unit" and "… module" described in the specification mean units that process at least one function or operation, and these may be implemented by hardware, software, or a combination of hardware and software.

[0011] Furthermore, terms including ordinal numbers such as "first" or "second" used in this specification may be used to describe various components, but the above components should not be limited by the above terms. The above terms may be used for the purpose of distinguishing one component from another.

[0012] Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. However, the embodiments can be implemented in various different forms and are not limited to the examples described herein.

[0013] This application is a continuation application of Korean Patent Application No. 10-2020-0185909. Therefore, the content described herein is based on the content of Korean Patent Application No. 10-2020-0185909. Accordingly, the content described in Korean Patent Application No. 10-2020-0185909 can be referred to for understanding the invention described herein, and even if it is omitted below, the content described in Korean Patent Application No. 10-2020-0185909 can be adopted for the invention described herein.

[0014] Hereinafter, embodiments will be described in detail with reference to the drawings.

[0015] FIG. 1 is a diagram for explaining an example of classifying input data into at least one class.

[0016] In FIG. 1, an example of input data 110, a classification model 120, and a classification result 130 is shown. FIG. 1 shows that the input data 110 is classified into a total of three classes, but the number of classes is not limited to the example of FIG. 1.

[0017] The type of the input data 110 is not limited. For example, the input data 110 may correspond to various types of data such as images, texts, and audios.

[0018] The classification model 120 may classify the input data 110 into a specific class. For example, the classification model 120 may calculate the probability that the input data is classified for each class using a Softmax function and cross entropy.

[0019] For example, assuming that the input data 110 is an image, and the first class is male and the second class is female, the classification model 120 will classify the input image into either the first or second class. Even if the input data 110 is an animal image, the classification model 120 will classify the input image into the class that is deemed more similar to the first or second class.

[0020] On the other hand, the classification model 120 may be trained using training data. In this case, the distribution of the training data can influence the training of the classification model 120. In other words, the performance of the classification model 120 may depend on the distribution of the training data.

[0021] The classification model 120 may be trained in a way that reduces the error between the output from the classification model 120 and the actual correct answer. For example, if the training data represents a long-tail distribution (for example, if some classes have many samples and other classes have very few), then when cross-entropy based on the softmax function is used to train the classification model 120, there is a problem in that the classification model 120 is trained to overfit to the major class.

[0022] To address the overfitting problem described above, conventional methods have involved undersampling data in the major class of the training data or oversampling data in the minor class. However, these methods assume that the training data is uniformly distributed. If the training data is not actually uniformly distributed, using these methods will result in poor training performance for the classification model 120.

[0023] The specific details of how the distribution of training data affects the training of classification model 120 will be discussed later, with reference to Figures 2 and 3.

[0024] Figures 2A and 2B illustrate an example of how a classification model works during the learning and inference phases.

[0025] Figure 2A shows an overview of the learning stage 210 and inference stage 220 based on the conventional classification models 213 and 223.

[0026] As described above with reference to Figure 1, the learning result (or conditional probability) of the classification model 213 may be determined by the label distribution 212 of the training data 211. In other words, the learning result of the classification model 213 may change depending on the label distribution 212 of the training data. The relationship between the label distribution 212 of the training data and the learning result of the classification model 213 can be explained by the following equation 1. Here, equation 1 is the equation corresponding to Bayes' rule. [Equation 1]

number

[0027] In equation 1, p s (y|x) represents the conditional probability that given the input training data x, it will be classified into class (label) y, and p s (y) represents the probability that class y appears (i.e., the distribution of class y). Also, in equation 1, p s (x|y) represents the probability that x is training data given class y, and p s (x) represents the probability of the training data appearing (i.e., the distribution of training data x). Also, in equation 1, c represents an element included in class y. In other words, class y is c is a variable representing each class.

[0028] Referring to Equation 1, the learning result of classification model 213 is p s(y|x) is the distribution of class y, p s It can be seen that it is related to (y). In other words, the learning results of classification model 120 are entangled with the distribution of class y. Therefore, the learning results of classification model 120 depend on the distribution of class y in the training data.

[0029] On the other hand, the label distribution 222 of the input data 221 in the inference stage 220 may differ from the label distribution 212 of the training data. For example, as shown in Figure 2A, the label distribution 212 of the training data and the label distribution 222 of the input data do not have to be equal, and moreover, the label distribution 212 of the training data and the label distribution 222 of the input data can represent completely opposite trends.

[0030] If the label distribution 212 of the training data and the label distribution 222 of the input data are not equal, the classification result of the input data 221 may not be accurate in the inference stage 220. As described above with reference to Equation 1, the label distribution 212 of the training data and the training result p s Because (y|x) are entangled with each other, if the label distributions 222 of the input data and 212 of the training data are different, a discrepancy may occur between the label distribution 222 of the input data and the conditional probability of the trained classification model 223. Therefore, the greater the discrepancy between the label distribution 212 of the training data and 222 of the input data, the lower the accuracy of the classification result by the trained classification model 223 may be. This is a major factor that degrades the performance of the trained classification model 223.

[0031] Figure 2B shows an overview of the learning stage 230 and inference stage 240 of classification models 233 and 243 according to one embodiment.

[0032] As described above with reference to Figure 2A, according to the conventional technique, the learning results of the classification model 213 are entangled with the label distribution 212 of the training data 211.

[0033] Therefore, in the learning stage 230 according to one embodiment, in the first mathematical formula (e.g., Formula 1) corresponding to the classification model 233, so that the label distribution 232 of the source data 231 does not affect the learning of the classification model 233, learning is performed based on the second mathematical formula in which the components corresponding to the label distribution 232 of the source data 231 are disentangled. In various embodiments, the classification model can be learned using only the conditional distribution of the sample x ( JPEG2026113586000003.jpg6150) and the distribution of the sample x in the training data ( JPEG2026113586000004.jpg6150) when the label y is given. Also, in the inference stage 240, the components related to the label distribution 242 of the target data 241 are applied to the learned classification model 243, and based on this, the classification of the target data 241 proceeds. In various embodiments, the components related to the label distribution may be determined using various methods such as Monte Carlo approximation formulas, but are not limited thereto. Thereby, regardless of whether the label distribution 232 of the source data and the label distribution 242 of the target data are the same, the target data 241 can be accurately classified.

[0034] FIG. 3 is a diagram for explaining an example of a classification result according to the distribution of source data for learning.

[0035] In FIG. 3, a graph showing the correlation between ps(y) representing the distribution of class y by the source data, p t (y) representing the distribution of class y by the target data, and the classification result (Avg.Prob.) of the target data by the learned classification model 223 is shown.

[0036] Referring to FIG. 3, the classification result (Avg.Prob.) of the target data is derived in the same way as p s (y). This is because the learning of the classification model 213 is performed depending on p s (y). Therefore, as described above with reference to FIG. 2A, the classification result of the target data by the learned classification model 223 is pt Although it should be expressed similarly to (y), the actual classification result of the target data (Avg. Prob.) is p t (y) may not match.

[0037] According to the data classification device of one embodiment, the trained classification model 243 can accurately classify the target data 241 regardless of the distribution 232 of the source data. In other words, the data classification device of one embodiment can accurately classify the target data 221 regardless of whether the distribution 232 of the source data and the distribution 242 of the target data are different. Specifically, the data classification device of one embodiment uses equation 1, p s (y|x) from p s The result of disentanglement of (y) ( By training on JPEG2026113586000005.jpg10150) and incorporating information representing the label distribution of the target data into the training results, the system can ultimately operate in a way that accurately classifies the target data 221.

[0038] Figure 4 is a flowchart showing an example of a method for classifying data related to one embodiment.

[0039] The data classification method shown in Figure 4 can be performed by a data classification device described later, referring to Figure 5. Specifically, the data classification method shown in Figure 4 can be performed by the processor shown in Figure 5. Therefore, it is understandable to any ordinary engineer that the processor shown in Figure 5 is the entity that executes the steps included in the flowchart in Figure 4.

[0040] In step 410, the processor learns a classification model to generate a first output value from a second formula obtained by untangling a component corresponding to the label distribution of the source data from a first formula corresponding to a classification model that classifies the input data into at least one class. In various embodiments, the classification model is determined by the conditional distribution of the sample x given a label y ( JPEG2026113586000006.jpg6150) and the distribution of sample x of the training data ( Training can be performed using only JPEG2026113586000007.jpg6150).

[0041] Here, the first equation represents Bayes' rule as expressed in Equation 1. That is, the first equation represents the probability that the input data is classified into at least one class. The component corresponding to the label distribution of the source data is p in Equation 1. s (y) is also acceptable.

[0042] As described above with reference to Figures 2 and 3, according to Equation 1, the label distribution of the source data (i.e., the distribution of at least one class based on the source data) and the learning result of the classification model may be entangled with each other. Therefore, the processor can resolve the entanglement between the label distribution of the source data and the learning result of the classification model by learning the result of unentangling the component corresponding to the label distribution of the source data in Equation 1.

[0043] Specifically, the processor may perform the following two steps using Equation 1. As the first step, the processor performs p according to Equation 1. s (y|x) from p s Untangle (y). That is, the processor untangles equation 1 JPEG2026113586000008.jpg10150 to p s Untangle (y) to obtain the second equation ( The second step is to generate JPEG2026113586000009.jpg10150). In the second step, the processor calculates the second formula ( JPEG2026113586000010.jpg10150) s (x) p s (y) has a uniform distribution of p u Replace (y) with p. u (y=c)=1 / C means that C is the total number of classes.

[0044] Through the two steps described above, the processor can learn a second equation such that equation 2 holds true. [Equation 2]

number

[0045] In equation 2, JPEG2026113586000012.jpg6150 is the first output value resulting from the training of the classification model, and represents the logit of class y of the classification model. In other words, the processor uses the first output value ( The second formula ( JPEG2026113586000013.jpg6150) is generated. Learn about JPEG2026113586000014.jpg10150). Here, the second formula ( Learning JPEG2026113586000015.jpg10150) is the log term of equation 2. This means performing calculations on JPEG2026113586000016.jpg10150.

[0046] For example, the processor is the second formula ( At least one approximation formula for JPEG2026113586000017.jpg10150) and information representing the label distribution of the source data (p s Using (y), the first output value ( The second formula ( JPEG2026113586000018.jpg6150) is generated. JPEG2026113586000019.jpg10150) may be learned. Here, at least one approximation formula includes a regularized Donsker-Varadhan representation and a Monte Carlo approximation formula.

[0047] Specifically, the normalized DV expression can be represented by the following equation 3. The normalized DV expression obtained by equation 3 acts to enable the learning of the log term in equation 2. [Equation 3]

number

[0048] In equation 3, JPEG2026113586000021.jpg6150 and JPEG2026113586000022.jpg6150 is, This means any distribution that satisfies JPEG2026113586000023.jpg6150. Also, in equation 3, some domains All functions of JPEG2026113586000024.jpg6150 For JPEG2026113586000025.jpg6150, the function T that minimizes the normalized DV expression is: JPEG2026113586000026.jpg6150 and Assume that this is the log-likelihood ratio of JPEG2026113586000027.jpg6150. Also, equation 3 is given for any expectation when the expected values ​​are finite. This applies to JPEG2026113586000028.jpg6150.

[0049] The processor is JPEG2026113586000029.jpg6150 and Substitute JPEG2026113586000030.jpg6150 into equation 3 and parameterize it using the logit of a deep neural network. Select the set of functions for JPEG2026113586000031.jpg6150. This will result in formula 2 JPEG2026113586000032.jpg6150 can access the target objective of Equation 3. In other words, the processor is the log term of Equation 2. JPEG2026113586000033.jpg10150 (that is, the optimal The image JPEG2026113586000034.jpg6150) may be trained using the following equation 4. For example, the processor can continue training until an equality is established between the left and right sides of the following equation 4. [Equation 4]

number

[0050] On the other hand, through equation 4, p u (x|y) and p u It is difficult to accurately estimate the expected value of (x). Therefore, the processor can substitute the Monte Carlo approximation formula into equation 4 and proceed with learning, and the Monte Carlo approximation formulas are given by equations 5 and 6 below. [Formula 5]

number

number

[0051] In equations 5 and 6 described above, JPEG2026113586000038.jpg6150 and JPEG2026113586000039.jpg6150 represents the i-th sample and label, respectively. Also, JPEG2026113586000040.jpg6150 represents the total number of samples. JPEG2026113586000041.jpg6150 represents the number of samples in Class C.

[0052] In equation 5, the sample and label pair The importance sampling for JPEG2026113586000042.jpg6150 is p s Using the sample (x), p u This is used to approximate the expected value for (x), which is given by the following equation 7. [Equation 7]

number

[0053] In equation 7, Regarding JPEG2026113586000044.jpg6150, Let's assume the filename is JPEG2026113586000045.jpg6150.

[0054] Alternatively, the processor may learn a classification model using information that represents the normalization of the label distribution of the source data.

[0055] The processor applies equations 5 and 6 to equation 4 to create a new loss that normalizes the logit to approach equation 2. The image JPEG2026113586000046.jpg6150 can be processed using the following equations 8 and 9. [Equation 8]

number

number

[0056] Equations 8 and 9 are, Sample and label pair, JPEG2026113586000049.jpg6150 This is defined for a single batch of JPEG2026113586000050.jpg6150. Also, in equations 8 and 9, JPEG2026113586000051.jpg6150 represents non-negative hyperparameters, and C represents the total number of classes. Also, JPEG2026113586000052.jpg6150 represents the number of samples of class c. JPEG2026113586000053.jpg6150 represents a set of classes that exist within a batch.

[0057] On the other hand, normalizing major classes more strongly than minor classes is even more desirable to improve the performance of the classification model. Therefore, the processor, JPEG2026113586000054.jpg6150 is class c of formula 8 and This may be applied as a weight for normalizing JPEG2026113586000055.jpg6150. The weight may be based on the size of each class.

[0058] In step 420, the processor generates a second output value by reflecting information representing the label distribution of the target data in the first output value.

[0059] For example, the processor may reflect information representing the label distribution of the target data in the first output value by multiplying it. The second output value generated by the processor reflecting information representing the label distribution of the target data in the first output value can be expressed as shown in the following equation 10. [Equation 10]

number

[0060] In equation 10, x represents the target data, and y represents the class (label). Also, JPEG2026113586000057.jpg6150 is the second output value, representing the classification model when input x is the input data. This represents the probability that JPEG2026113586000058.jpg6150 is classified into class (label) y.

[0061] In other words, the processor, through step 410, generates the first output value ( JPEG2026113586000059.jpg6150) contains information representing the label distribution of the target data ( By reflecting JPEG2026113586000060.jpg6150), a second output value represented by formula 10 can be generated.

[0062] In step 430, the processor uses the second output value to classify the target data into at least one class.

[0063] For example, the processor can use equation 10 to generate output data in which the target data is classified into at least one class. In this case, the final loss function based on cross-entropy for the output data can be defined as shown in equation 11 below. [Equation 11]

number

[0064] In equation 11, JPEG2026113586000062.jpg6150 represents the final loss function. JPEG2026113586000063.jpg6150 is, This refers to a non-negative hyperparameter that determines the normalized intensity of JPEG2026113586000064.jpg6150. Also, in equation 11, The image JPEG2026113586000065.jpg6150 can be calculated using the following formula 12. [Equation 12]

number

[0065] On the other hand, cross-entropy loss ( JPEG2026113586000067.jpg6150) is represented by the following equation 13. [Equation 13]

number

[0066] The processor can derive equation 12 based on equation 13. The processor, Based on JPEG2026113586000069.jpg6150, the final loss function ( It can perform calculations on JPEG2026113586000070.jpg6150.

[0067] Figure 5 is a configuration diagram showing an example of a data classification device according to one embodiment.

[0068] The data classification device 500 shown in Figure 5 performs the data classification method described above by referring to Figure 4. Therefore, even if omitted below, it is easy for an ordinary engineer to understand that the above-described method for classifying data by referring to Figure 4 can also be applied to the data classification device 500.

[0069] The data classification device 500 includes a memory 510 and a processor 520.

[0070] Memory 510 is operablely connected to processor 520 and stores at least one program for processor 520 to operate. Memory 510 also stores all data related to the above, such as training data, input data, and class information, as shown in Figures 1 to 4.

[0071] For example, memory 510 can temporarily or permanently store data processed by processor 520. Memory 510 may include, but is not limited to, magnetic storage media or flash storage media. Memory 510 may include internal memory and / or external memory, and may include volatile memory such as DRAM, SRAM, or SDRAM; non-volatile memory such as one-time programmable ROM (OTPROM), PROM, EPROM, EEPROM, mask ROM, flash ROM, NAND flash memory, or NOR flash memory; flash drives such as SSDs, compact flash (CF) cards, SD cards, Micro-SD cards, Mini-SD cards, Xd cards, or memory sticks; or storage devices such as HDDs.

[0072] The processor 520 executes the data classification method described above, referring to Figures 1 to 4, according to the program stored in memory 510.

[0073] The processor 520 learns a classification model to generate a first output value from a second equation obtained by untangling a component corresponding to the label distribution of the source data from a first equation corresponding to a classification model that classifies the input data into at least one class. Here, the first equation corresponds to Bayes' rule, which represents the probability that the input data is classified into at least one class.

[0074] For example, the processor 520 learns a classification model using at least one approximation formula for the second mathematical formula and information representing the label distribution of the source data. Here, at least one approximation formula includes a regularized Donsker-Varadhan representation and a Monte Carlo approximation formula.

[0075] Furthermore, processor 520 learns a classification model using information representing the normalization of the label distribution of the source data.

[0076] Furthermore, the processor 520 generates a second output value by reflecting information representing the label distribution of the target data in the first output value. For example, the processor 520 may reflect the information representing the label distribution of the target data in the first output value by multiplying it.

[0077] Furthermore, the processor 520 uses the second output value to classify the target data into at least one class.

[0078] For example, processor 520 may mean a data processing device embedded in hardware having a physically structured circuit for executing functions expressed by code or instructions contained within a program. Here, examples of data processing devices embedded in hardware may include, but are not limited to, microprocessors, central processing units (CPUs), processor cores, multiprocessors, application-specific integrated circuits (ASICs), and field programmable gate arrays (FPGAs).

[0079] Figure 6 illustrates an example in which the second output value according to one embodiment is utilized.

[0080] Figure 6 shows a network configuration diagram including a server 610 and multiple terminals 621 to 624 according to one embodiment.

[0081] Server 610 may be an intermediary device connecting multiple terminals 621-624. Server 610 may provide an intermediary service that enables data to be sent and received between multiple terminals 621-624. Server 610 and the multiple terminals 621-624 may be connected to each other via a communication network. Server 610 can transmit and receive data to and from multiple terminals 621-624 via the communication network.

[0082] Here, the communication network can be implemented as one of the following: a wired communication network, a wireless communication network, or a hybrid communication network. For example, the communication network may include mobile communication networks such as 3G, LTE (Long Term Evolution), LTE-A, and 5G. Furthermore, the communication network may include wired or wireless communication networks such as Wi-Fi, UMTS (Universal Mobile Telecommunisations System) / GPRS (General Packet Radio Service), or Ethernet.

[0083] The communication network may include local area networks such as Magnetic Secure Transmission (MST), Radio Frequency Identification (RFID), Near Field Communication (NFC), ZigBee, Z-Wave, Bluetooth, Bluetooth Low Energy (BLE), or Infrared Communication (IR). The communication network may also include local area networks (LAN), metropolitan area networks (MAN), or wide area networks (WAN).

[0084] Each of the multiple terminals 621-624 may be implemented as one of the following: a desktop computer, a laptop computer, a smartphone, a smart tablet, a smartwatch, a mobile terminal, a digital camera, a wearable device, or a portable electronic device. Furthermore, the multiple terminals 621-624 may run programs or applications.

[0085] For example, multiple terminals 621-624 may be running an application that can provide a mediation service. Here, mediation service means that users of multiple terminals 621-624 make video calls and / or voice calls to each other.

[0086] To provide the intermediary service, the server 610 may perform various classification tasks. For example, the server 610 may classify users into predetermined classes based on information provided by users of multiple terminals 621 to 624. In particular, if the server 610 receives facial images from people who have subscribed to the intermediary service (i.e., users of multiple terminals 621 to 624), the server 610 can classify the facial images into predetermined classes for various purposes. For example, the predetermined classes may be based on a person's gender, or on a person's age.

[0087] At this time, the second output value generated by the method described above (refer to Figure 4) is stored in the server 610, and the server 610 can accurately classify the facial images into predetermined classes.

[0088] As described above, it is possible to generate a classification model that can accurately classify input data into predetermined classes, regardless of the distribution of the input data.

[0089] On the other hand, the method described above can be created with a program that can be executed by a computer and can be implemented on a general-purpose digital computer that runs the above program using a computer-readable recording medium. Furthermore, the data structure used in the method described above can be recorded on a computer-readable recording medium by various means. The above computer-readable recording mediums include storage media such as magnetic storage media (e.g., ROM, RAM, USB, floppy disks, hard disks, etc.) and optical reading media (e.g., CD-ROM, DVD, etc.).

[0090] A person with ordinary skill in the art related to this embodiment will understand that it can be implemented in a form that is modified without departing from the essential properties of the above-mentioned substrate. Accordingly, the disclosed method should be considered in an explanatory rather than restrictive sense, and the scope of rights should be interpreted as being as set forth in the claims and equivalent extent, rather than in the above description, and should be interpreted as including all differences within that scope.

Claims

1. The steps include: training the classification model to generate a first output value using a second formula obtained by disentangling a component corresponding to the label distribution of the source data from a first formula corresponding to a classification model that classifies the input data into at least one class; The steps include generating a second output value by reflecting information representing the label distribution of the target data in the first output value, A method for classifying data, comprising the step of classifying the target data into the at least one class using the second output value.

2. The method according to claim 1, wherein the first formula includes a formula corresponding to a Bayesian rule that represents the probability that the input data is classified into the at least one class.

3. The step of generating the second output value is: The method according to claim 1, wherein the first output value is reflected by multiplying it with information representing the label distribution of the target data.

4. The step of training the aforementioned classification model is: The method according to claim 1, wherein the classification model is trained using at least one approximation formula relating to the second formula and information representing the label distribution of the source data.

5. The method according to claim 4, wherein the at least one approximation formula includes a normalized Donsker-Varadan representation and a Monte Carlo approximation formula.

6. The step of training the aforementioned classification model is: The method according to claim 1, wherein the classification model is trained using information representing the normalization of the label distribution of the source data.

7. The step of training the aforementioned classification model is: The method according to claim 1, comprising training a classification model using only the distribution of sample x (p_s(x)) and the conditional distribution of sample x for label y (p_s(x|y)) from training data.

8. A computer-readable recording medium that stores a program for executing the method according to claim 1 on a computer.

9. Memory to store at least one program, A data classification device comprising a processor that operates by executing at least one of the aforementioned programs, The aforementioned processor, An apparatus for training a classification model such that a first output value is generated by a second mathematical formula obtained by untangling a component corresponding to the label distribution of source data from a first mathematical formula corresponding to a classification model that classifies input data into at least one class; generating a second output value by reflecting information representing the label distribution of target data in the first output value; and classifying the target data into the at least one class using the second output value.

10. The apparatus according to claim 9, wherein the first formula includes a formula corresponding to a Bayesian rule that represents the probability that the input data is classified into the at least one class.

11. The aforementioned processor, The apparatus according to claim 9, wherein the first output value is reflected by multiplying it with information representing the label distribution of the target data.

12. The aforementioned processor, The apparatus according to claim 9, wherein the classification model is trained using at least one approximation formula relating to the second mathematical formula and information representing the label distribution of the source data.

13. The apparatus according to claim 12, wherein the at least one approximation formula includes a normalized Donsker-Varadan representation and a Monte Carlo approximation formula.

14. The aforementioned processor, The apparatus according to claim 9, wherein the classification model is trained using information representing the normalization of the label distribution of the source data.

15. The aforementioned processor, The apparatus according to claim 9, which trains a classification model using only the distribution of sample x (p_s(x)) and the conditional distribution of the label y (p_s(x|y)) of sample x from the training data.

Images