KR20210102860A - Apparatus and method for processing imbalance data - Google Patents

Abstract

The present invention relates to a device and method for processing imbalanced data, wherein the device for processing the imbalanced data may comprise: an over-sampling part that acquires at least one data, and adds the at least one piece of data to a minority sample by further generating the virtual data; and an under-sampling part that detects data adjacent to the data of the minority sample from among the data of a plurality of samples, and removes the adjacent data from the plurality of samples. Therefore, the present invention is capable of preventing a degradation in performance of the learning model.

Description

Unbalanced data processing apparatus and method {APPARATUS AND METHOD FOR PROCESSING IMBALANCE DATA}

It relates to an apparatus and method for processing unbalanced data.

With the recent development of hardware technology related to information processing, it is possible to quickly process a large amount of data. This is further accelerating the development, research, introduction and use of machine learning technology that performs learning based on a large amount of data and acquires various information or performs computation or control processing based on rules obtained as a result of learning. Such machine learning technology is being used in various fields. For example, data search, forecasting of future data such as weather, speech recognition and text conversion, character recognition, language translation, vehicle autonomous driving, consumer tendency analysis and related services, games, smart grid and Internet of Things, etc. Artificial intelligence technology is being developed, adopted and used in various fields. Among machine learning, supervised learning is a method of learning using data and corresponding correct answers. It is divided into a classification method that classifies data into any one class and a regression method that infers a function consisting of a sequence of data. However, the classification method of supervised learning has a problem in that its performance is degraded when data is imbalanced. Specifically, data corresponding to one or more specific classes (multiple classes) are input in relatively large numbers, while data corresponding to other one or more classes (minority classes) are input in a relatively small number. However, if data to be classified is input later, there is a high possibility that an error will occur in their classification. That is, when data corresponding to a minority class is input as classification target data, the learning model is likely to misclassify these data into a majority class. Therefore, such data imbalance between classes has a problem in that the prediction accuracy of the learning model is improved while the recall is reduced, thereby degrading the overall performance of the learning model.

An object to be solved is to provide an imbalanced data processing apparatus and method that can solve the performance degradation of the learning model due to data imbalance between classes.

In order to solve the above problems, an apparatus and method for processing unbalanced data are provided as follows.

The imbalance data processing apparatus receives at least one data belonging to a fractional sample, an oversampling unit that further generates virtual data and adds it to the fractional sample, and at least one data belonging to the plurality of samples, the plurality of samples and an under-sampling unit configured to detect data adjacent to the data of the minority sample from among the data of , and remove the adjacent data from the plurality of samples.

The method for processing unbalanced data includes: acquiring at least one data, wherein the at least one data belongs to at least one of a minority sample or a multiple sample; generating virtual data further and adding it to the minority sample to perform oversampling and detecting data adjacent to the data of the minority sample from among the data of the plurality of samples, and performing undersampling by removing the adjacent data from the plurality of samples.

According to the above-described apparatus and method for processing unbalanced data, it is possible to solve various problems caused by data imbalance between classes.

According to the above-described unbalanced data processing apparatus and method, it is possible to prevent the performance degradation of the learning model for supervised learning, etc. even in a situation where one or more multiple classes with relatively large data and one or more minority classes with relatively small data exist. do.

According to the above-described imbalance data processing apparatus and method, there is an advantage in that the accuracy of classification of time series data, sound data, pattern data, etc. can be increased, and a recall rate can also be increased.

1 is a block diagram of an embodiment of an apparatus for processing unbalanced data.
2 is a block diagram of one embodiment of data.
3 is a graph showing a t-distribution stochastic embedding result for data before sampling processing.
4 is a graph showing a t-distribution stochastic embedding result for data after sampling processing.
5 is a block diagram of an embodiment of a method for processing unbalanced data.

In the following specification, the same reference numerals refer to the same elements unless otherwise specified. The term added with 'part' used below may be implemented in software or hardware, and one 'unit' may be implemented as one physical or logical part, or a plurality of parts may be implemented according to the consideration or arbitrary selection of a designer or user It is also possible that 'unit' is implemented as one physical or logical part, or one 'unit' is implemented with a plurality of physical or logical parts. Throughout the specification, when it is said that a part is connected to another part, this may mean a physical connection or an electrically connected part depending on the part and the other part. In addition, when it is said that a part includes another part, it does not exclude another part other than the other part unless otherwise stated, and it means that another part may be further included according to the designer's choice. do. Terms such as first and second are used to distinguish one part from another, and unless otherwise specified, they do not mean sequential expressions. Also, singular expressions may include plural expressions unless the context clearly dictates otherwise.

Hereinafter, an embodiment of an apparatus for processing unbalanced data will be described with reference to FIGS. 1 to 4 .

1 is a block diagram of an embodiment of an apparatus for processing unbalanced data, and FIG. 2 is a block diagram of an embodiment of data.

1 and 2 , the imbalance data processing apparatus 100 includes, in an embodiment, a data acquisition unit 101 that acquires at least one data 50 from the outside, and a data acquisition unit ( Based on the data 50 received from 101), the processor 110 capable of processing imbalanced data, training the learning model, and/or obtaining the trained learning model-based result, and the information obtained by the processor 110 It may include a data output unit 103 that can be output visually or audibly to the outside. Here, the data acquisition unit 101, the data output unit 103, and the processor 110 may transmit commands/instructions or data to one or both sides, for example, among cables, wireless communication networks, and circuit lines. It may be provided to sequentially or non-sequentially transmit a command/instruction or data through at least one.

The data acquisition unit 101 may acquire at least one data 50 according to a user's manipulation or reception from a memory device or other information processing device. The data acquisition unit 101 may include, according to an embodiment, a keyboard, a mouse, a tablet, a touch screen, a touch pad, a track ball, a track pad, a scanner device, an image capturing module, a motion detection sensor, a vibration sensor, a light receiving sensor, a pressure sensor, Proximity sensor, ultrasound scanner, microphone, data input/output terminal and/or communication module (eg, LAN card, short-distance communication module or mobile communication module, etc.) capable of receiving data from other external devices (eg, portable memory device, etc.) and the like.

The data 50 acquired by the data acquisition unit 101 is, for example, as shown in FIG. 2 , at least one data 1 to 4 belonging to any one sample 51 (hereinafter referred to as a prime sample) and/or Alternatively, at least one piece of data 5 to 13 belonging to another sample 52 (hereinafter, multiple samples) may be included. In this case, the minority sample 51 includes a relatively small number of data 1 to 4 compared to the majority sample 52 , and on the contrary, the majority sample 52 includes a relatively large number of data compared to the minority sample 51 . data 5 to 13 may be included. Depending on circumstances, the minority sample 51 may include very little data compared to the majority sample 52 . The data 1 to 4 of the minority sample 51 are data that can be classified into any one class (hereinafter referred to as a minority class) through supervised learning, etc., and the data 5 to 13 of the multiple sample 52 are different Data that can be classified into one class (hereinafter, multiple classes). That is, the minority sample 51 corresponds to the minority class, and the majority sample 52 corresponds to the majority class. In this case, if the number of data 1 to 4 in the minority sample 51 is significantly smaller than the number of data 5 to 13 in the majority sample 52, both samples 51 and 52 data 1 to 4 , 5 to 13), there is a problem in that the recall decreases even if the accuracy is increased. Although only two samples 51 and 52 are illustrated in FIG. 2 , this is exemplary, and the number of samples 51 and 52 or the number of data 1 to 13 is not limited thereto. According to an embodiment, the data acquisition unit 101 may acquire data belonging to three or more samples. Here, at least one sample of the three or more samples may include a relatively small number of data compared to the other two or more samples, and the at least one other sample may include a relatively large number of data with respect to the other two or more samples. In addition, according to an embodiment, the data 1 to 13 may include features processed by the pre-processing unit 111 to be described later or extracted by the feature extracting unit 112 .

The data 1 to 13 may include at least one pattern data, and the pattern data may include time series data that changes over time. The pattern data may include, for example, a sine wave or a harmonic wave. According to an embodiment, the data 1 to 13 may include sound data such as auscultation sound data (eg, heart sound), voice, ambient sound, noise, or music. Here, the heart sound is a beating sound of the heart, and it is possible to easily measure whether the heart is abnormal, such as a valve disease, based on whether the heart sound is normal. These heart sounds are time-series pattern data. The data 1 to 13 acquired by the data acquisition unit 101 may be immediately or sequentially transmitted to the processor 110 . Before delivery to the processor 110 , the data 1 to 13 may be temporarily or non-temporarily stored in a storage unit (not shown). Here, the storage unit may include at least one of a main memory device and an auxiliary memory device.

The processor 110 may appropriately process unbalanced data by acquiring at least one data 1 to 13 and performing oversampling and undersampling on the at least one data 1 to 13 .

1 , the processor 110 may include a pre-processing unit 111 , a feature extraction unit 112 , a sampling processing unit 113 , an imbalance checking unit 114 , and a classification processing unit 115 . can According to an embodiment, at least one of the pre-processing unit 111 , the feature extraction unit 112 , the imbalance checking unit 114 , and the classification processing unit 115 may be omitted. Here, the preprocessing unit 111 , the feature extracting unit 112 , the sampling processing unit 113 , the imbalance checking unit 114 , and the classification processing unit 115 may be logically separated or physically separated.

The preprocessor 111 may convert the data 1 to 13 into a state in which processing is possible or easy. For example, when the data 1 to 13 are frequency-based data, the preprocessor 111 may perform preprocessing of the data 1 to 13 by applying a Fourier transform to the data 1 to 13 . .

The feature extractor 112 may extract at least one feature from the data 1 to 13 , for example, sound data. The features may include features corresponding to the minority sample 51 and features corresponding to the majority sample 52 . In this case, the feature extraction unit 112 may extract a feature based on any one of time, frequency, and statistics, or may extract a feature based on a plurality of them and then fuse them. The feature extraction unit 112 may include, for example, Mel-frequency cepstral coefficients (MFCC), Mel spectrogram, zero crossing rate, chroma, and contrast. The feature may be extracted using one of (contrast) methods or a combination of two or more of them. Here, the Mel spectrogram divides the input signal into frames, applies a Fourier transform to it, obtains a spectrum, and then filters it through a Mel filter to express the data in the frequency domain. It is a vector obtained through cepstral analysis for gram. The zero crossing rate is a method of using a value obtained by measuring the number of times a value of a voice signal or time series data changes to 0 as a value for a feature. Chroma is a vector that represents pitch as a pitch class representing 12 scales. Contrast is a method of classifying music genres through characteristics while indicating characteristics of an octave-based spectral band.

The sampling processing unit 113 may perform processing on the data 1 to 13 , such as adding new data to the given data 1 to 13 or removing some of the given data 1 to 13 . Here, the given data 1 to 13 may be data 1 to 13 acquired by the data acquisition unit 101, data pre-processed by the pre-processing unit 111, and/or feature extraction unit It may be a feature corresponding to the data 1 to 13 extracted in 112 .

According to an embodiment, the sampling processing unit 113 may include an oversampling processing unit 113a performing oversampling processing and an undersampling processing unit 113b performing undersampling processing. The oversampling processing unit 113a may further generate virtual data and add it to the decimal sample 51 . That is, the decimal sample 51 further includes other virtual data (or features) in addition to the existing data (1 to 4, which may include features as described above). In this case, the oversampling processing unit 113a selects at least one (eg, central data) from one or more data 1 to 4 of the decimal sample 51 , and connects the selected data and other data with a virtual line segment. Then, at least one virtual data may be generated by arbitrarily adding at least one virtual data on the virtual line segment. Accordingly, the number of data 1 to 4 corresponding to the decimal sample 51 increases. The under-sampling processing unit 113b may perform under-sampling by deleting at least one data from among the data 5 to 13 of the plurality of samples 52 . In this case, the undersampling processing unit 113b performs undersampling by using the distance between the data (1 to 13, which may include features belonging to the plurality of samples 52 as described above) in the samples 51 and 52 . can do. For example, the under-sampling processing unit 113b may perform at least one of data (1 to 4, or extracted features) of the minority sample 51 among data (5 to 13, or extracted features) of the multiple samples 52 and At least one data adjacent to each other is obtained, and at least one data adjacent to at least one of the data 1 to 4 of the minority sample 51 is removed from the majority sample 52 to remove the data 5 to 13 of the majority sample 51 ) can be reduced. According to an embodiment, the sampling processing unit 113 may perform at least one of the above-described oversampling and undersampling in response to a result of determining whether data imbalance has occurred by the imbalance checking unit 114 .

3 is a graph showing a t-distribution stochastic embedding result for data before sampling processing, and FIG. 4 is a graph diagram illustrating a t-distribution stochastic embedding result for data after sampling processing.

The imbalance check unit 114 is based on the data 1 to 13 acquired by the data acquisition unit 101, the data pre-processed by the pre-processing unit 111, and/or the features extracted by the feature extraction unit 112. Class imbalance can be determined. According to an embodiment, the imbalance check unit 114 includes data 1 to 4 of the minority sample 51 processed by the oversampling unit 113a and the multiple samples ( ) processed by the undersampling unit 113b. 52) by applying a t-distribution stochastic embedding (T-SNE: t-Stochastic Nearest Neighbor) to (5 to 13) to the data of (5) to determine whether the class imbalance. The t-distribution stochastic embedding is a machine learning method that reduces high-dimensional data to low-dimensional data, and reduces three-dimensional data to two-dimensional or three-dimensional data in the form of a graph as shown in FIGS. 3 and 4 . can be visualized. As illustrated in FIG. 3 , in the case of data on which oversampling or undersampling is not performed, data is mixed due to class imbalance. On the other hand, after oversampling and undersampling are performed by the oversampling unit 113a and the undersampling unit 113b, respectively, as shown in FIG. 4 , they are classified into a more improved form.

According to an embodiment, the classification processing unit 115 includes data 1 to 4 of the decimal sample 51 , virtual data added to the decimal sample 51 by the oversampling unit 113a, and undersampling. Class classification may be performed based on data (at least one of 5 to 13) of the plurality of samples 52 from which a part has been removed by the unit 113b. For this, the classification processing unit 115 may use at least one learning model. Here, the at least one learning model is, for example, a deep neural network (DNN), a convolutional neural network (CNN), a recurrent neural network (RNN), a deep trust neural network (DBN, Deep Belief Network, Deep Q-Networks, Long short term memory (LSTM), Random Forest, Multi-layer Perceptron, Support Vector Machine (SVM: support) vector machine), a generative adversarial network (GAN) and/or a conditional adversarial neural network (cGAN), etc. It may be implemented or implemented by combining two or more learning models. Meanwhile, the above-described learning models are exemplary, and in addition to these, the classification processing unit 115 may perform class classification based on at least one other learning model that a designer or a user can consider.

The above-described processor 100 is, for example, a central processing unit (CPU, Central Processing Unit), a micro controller unit (MCU, Micro Controller Unit), a microcomputer (Micom, Micro Processor), an application processor (AP, Application Processor) , an electronic control unit (ECU, Electronic Controlling Unit) and/or other electronic devices capable of processing various calculations and generating control signals. These devices may be implemented using, for example, one or two or more semiconductor chips. In addition, the processor 100 may execute a program stored in the storage unit to perform predefined operations, judgment, processing and/or control operations, etc., and the program stored in the storage unit may be directly created and It may be modified, it may be transmitted, installed, and updated from another memory device, and/or it may be obtained or updated through an electronic software distribution network accessible through a wired or wireless communication network.

The data output unit 103 visually or aurally outputs the processing result of the sampling processing unit 113, the graph related to the imbalance obtained by the imbalance checking unit 114, the classification processing result by the classification processing unit 115, and the like. may be output and/or transmitted to an external memory device or other information processing device. The data output unit 103 may include, for example, a display, a printer device, a speaker device, an image output terminal, a data input/output terminal, and/or a communication module, but is not limited thereto.

The above-described imbalanced data processing apparatus 100 may be implemented using one information processing apparatus or a combination of two or more information processing apparatuses. Here, the information processing device, according to an embodiment, is a computer device for a server, a desktop computer, a laptop computer, a smart phone, a tablet PC, a smart watch, a head mounted display (HMD) device, a navigation device, a portable game machine, and a personal device. Digital assistants (PDA: Personal Digital Assistant), digital televisions, electronic billboards, set-top boxes, home appliances (refrigerators and washing machines, etc.), artificial intelligence sound reproduction devices, manned vehicles (cars, buses, vehicles such as two-wheeled vehicles, etc.), unmanned It may include, but is not limited to, a moving object (robot cleaner, etc.), a construction machine, a manned vehicle, an unmanned aerial vehicle, a household or industrial robot, and/or an industrial machine. The one or more information processing devices may include various devices capable of performing training of a data-based learning model and/or information acquisition based on the trained learning model, in addition to those exemplified.

Hereinafter, an embodiment of a method for processing unbalanced data will be described with reference to FIG. 5 .

5 is a block diagram of an embodiment of a method for processing unbalanced data.

As shown in FIG. 5 , at least one data is first obtained, and the at least one data obtained may belong to any one of a plurality of samples ( 200 ). In this case, the plurality of samples may be divided into at least one minority sample or at least one multiple sample according to the number of data included. The at least one data may belong to at least one of at least one minority sample and at least one majority sample. According to an embodiment, the data may include sound data such as a heart sound or a feature extracted from the sound data. The features extracted from the acoustic data may be, for example, pre-processed using a Fourier transform on the acoustic data, and then extracted from the pre-processed data based on any one of time, frequency, and statistics. Extraction of features may be performed based on Mel frequency cepstral coefficients, Mel spectrogram, zero crossing, chroma and/or contrast, and the like.

If necessary, it may be checked whether the data is unbalanced ( 402 ). Whether there is an imbalance can be performed by applying t-distribution stochastic embedding to the data. Checking whether the data is unbalanced may be performed before the oversampling process 204 and the undersampling process 206 as shown in FIG. 5 , and at least one of the oversampling process 204 and the undersampling process 206 . may be performed simultaneously with or sequentially, and/or may be performed after the oversampling process 204 and the undersampling process 206 are processed. According to an embodiment, the determination of whether data is imbalanced may be performed on data before oversampling and undersampling, and also on data after oversampling and undersampling are processed.

Oversampling may be performed on the fractional sample (404). Oversampling can be performed by adding more hypothetical data to a fractional sample. According to an embodiment, the oversampling selects at least one of one or more data of a small sample, connects the selected data and other data with a virtual line segment, and then randomly adds at least one virtual data on the virtual line segment. It can also be done by

Undersampling may be performed on multiple samples following, preceding, or concurrently with oversampling (406). Undersampling may be performed in a manner that removes some of the data of multiple samples. Specifically, for example, the undersampling may be performed by selecting at least one data that is close to at least one data of a minority sample from among at least one data belonging to a plurality of samples, and deleting the selected at least one data from the plurality of samples. may be

After the oversampling 404 and the undersampling 406 are performed, class classification may be further performed based on the oversampled prime-sample data and the under-sampled multiple-sample data according to an embodiment ( 408 ). Class classification may be performed based on at least one learning model, for example, deep neural networks, convolutional neural networks, recurrent neural networks, deep trust neural networks, deep Q-networks, long and short-term memories, random forests, multi-layer perceptrons, support vectors. It may be performed based on a learning model of at least one of a machine, a generative adversarial neural network, and a conditional generative adversarial neural network. Class classification can be omitted.

The unbalanced data processing method according to the above-described embodiment may be implemented in the form of a program that can be driven by a computer device. Here, the program may include program instructions, data files, and data structures alone or in combination. The program may be designed and manufactured using machine code or high-level language code. The program may be specially designed to implement the above-described method, or may be implemented using various functions or definitions that are known and available to those skilled in the art of computer software. In addition, here, the computer device may be implemented including a processor or memory that enables the function of the program to be realized, and may further include a communication device if necessary. A program for implementing the above-described unbalanced data processing method may be recorded in a computer-readable recording medium. The computer-readable recording medium includes, for example, a solid state drive (SSD), a semiconductor storage device such as a ROM, RAM or flash memory, a magnetic disk storage medium such as a hard disk or a floppy disk, a compact disk or DVD, etc. It may include at least one type of physical device capable of storing a specific program executed in response to a call of a computer, such as an optical recording medium, a magneto-optical recording medium such as a floppy disk, and a magnetic tape.

Although an embodiment of the apparatus and method for processing unbalanced data has been described above, the apparatus and method for processing unbalanced data are not limited to the above-described embodiment. An apparatus or method that can be implemented by a person skilled in the art by modifying and modifying based on the above-described embodiment may also be an example of the above-described imbalanced data processing apparatus and method. For example, the described techniques are performed in an order different from the described method, and/or the described components of the system, structure, apparatus, circuit, function, component, etc., are combined or combined in a different manner than the described method. , may be an embodiment of the above-described unbalanced data processing apparatus and method even if substituted or substituted by other components or equivalents.

100: unbalanced data processing device 101: data acquisition unit
111: preprocessor 112: feature extraction unit
113: sampling processing unit 113a: oversampling processing unit
113b: under-sampling processing unit 114: imbalance check unit
115: classification processing unit

Claims (10)

an oversampling unit that receives at least one data belonging to a prime number sample, generates additional virtual data, and adds the data to the prime sample; and
An imbalance data processing apparatus comprising a; . According to claim 1,
The at least one data is an imbalance data processing apparatus including acoustic data. According to claim 1,
a preprocessor that performs a Fourier transform on the acoustic data; and
A feature extraction unit for extracting features from the sound data preprocessed by the preprocessor;
wherein the at least one data includes the extracted feature. According to claim 1,
An imbalance checking unit for confirming data imbalance based on t-distribution stochastic embedding for the at least one data or for the data of the prime sample and the plurality of samples processed by the over-sampling section and the under-sampling section; Unbalanced data processing unit further comprising. According to claim 1,
Class classification is performed based on at least one learning model on the data of the few samples and the plurality of samples processed by the oversampling unit and the undersampling unit, wherein the at least one learning model is a random forest (Random Forest). ), a multi-layer perceptron (MLP), a deep neural network (DNN), a convolutional neural network (CNN), and a recurrent neural network (RNN). Unbalanced data processing apparatus further comprising a processing unit. obtaining at least one data, wherein the at least one data belongs to at least one of a minority sample or a majority sample;
performing oversampling by further generating virtual data and adding it to the fractional sample; and
and detecting data adjacent to the data of the minority sample from among the data of the plurality of samples, and performing undersampling by removing the adjacent data from the plurality of samples. 7. The method of claim 6,
wherein the at least one data includes sound data. 8. The method of claim 7,
The at least one data is an imbalance data processing method including a feature extracted from sound data on which Fourier transform has been performed. 7. The method of claim 6,
Checking the data imbalance based on the t-distribution stochastic embedding for the at least one data or for the data of the fractional sample and the plurality of samples processed by the oversampling unit and the undersampling unit; How to handle unbalanced data. 6. The method of claim 5,
Class classification is performed based on at least one learning model on the data of the few samples and the plurality of samples processed by the oversampling unit and the undersampling unit, wherein the at least one learning model is a random forest (Random Forest). ), a multi-layer perceptron (MLP), a deep neural network (DNN), a convolutional neural network (CNN), and a recurrent neural network (RNN) comprising at least one of ; Unbalanced data processing method further comprising.

Images