Earticle

비만은 세계보건기구에서 질병으로 규정하고 있으며, 신체 내외부적인 영향을 나타낸다. 본 연구는 머신러닝을 이용 하여 생활패턴에 따른 비만도를 예측하고자 한다. 머신러닝에 이용한 데이터는 오픈데이터를 사용하였으며, 머신러닝 모델은 구글 코랩, 파이썬을 이용하고 모델 구성은 Light Gradient Boosting Machine, Extreme Gradient Boosting, Decision Tree Classifier, K Neighbors Classifier, Naive bayes 총 5개의 모델로 구성하였다. 각 모델 성능 평가 지표는 정확도, area under curve, 재현율, 정밀도, F1-score로 평가하였다. 해당 데이터를 활용한 5개 모델 중 Light Gradient Boosting Machine이 모든 지표에서 성능이 가장 우수 했으며, 지표에 대한 결과는 정확도 0.9601, area under curve 0.9981, 재현율 0.9601, 정밀도 0.9611, F1 score 0.9601이었다. 본 연구를 통해 기본적인 생활 패턴에 따른 비만도 예측을 통해 비만에 대한 사전 예방이 가능할 것으로 사료된다.

Obesity is defined as a disease by the World Health Organization and indicates internal and external influences on the body. This study aims to predict obesity according to lifestyle patterns using machine learning. The data used for machine learning used open data, and the machine learning model used Google Colab and Python. The model configuration consisted of a total of five models: Light Gradient Boosting Machine, Extreme Gradient Boosting, Decision Tree Classifier, K Neighbors Classifier, and Naive Bayes. The performance evaluation indices for each model were accuracy, area under curve, recall, precision, and F1-score. Among the five models using the data, Light Gradient Boosting Machine showed the best performance in all indices, and the results for the indices were accuracy 0.9601, area under curve 0.9981, recall 0.9601, precision 0.9611, and F1-score 0.9601. Through this study, it is believed that it will be possible to prevent obesity in advance by predicting obesity according to basic lifestyle patterns.

최근 침입 탐지 시스템에 대한 관심이 증대되고 있다. 침입 탐지 시스템에서 침입 여부 확인을 위하여 패턴매칭 기법이 주로 사용된다. 기존의 패턴매칭 기법들은 다양한 공격 패턴들에 대한 패턴 비교 시간이 많이 소요되는 문제점이 있었다. 본 논문에서는 기존의 패턴매 칭 기법들이 가지고 있는 문제점을 해결하기 위하여 새로운 침입 탐지 시스템을 제안한다. 제안한 시스템은 효율적인 패턴 비교를 위하여 룰 패턴을 분류한다. 분류된 패턴은 매칭을 위하여 정형화된 트리로 구현한다. 그러므로, 본 논문에서 제안한 침입 탐지 시스템 모델은 효율적으로 네트워크 침입 탐지를 수행할 수 있다.

Recently, lots of researchers work focused on the intrusion detection system. Pattern matching technique is commonly used to detect the intrusion in the system, However, the method requires a lot of time to match between systems rule and inputted packet data. This paper proposes a new intrusion detection system based on the pattern matching technique. Proposed system reduces the required time for pattern matching by using classified system rule. The classified rule is implemented with a general tree for efficient pattern matching. Thereby, proposed system could perform network intrusion detection efficiently.

본 논문은 서베일런스 네트워크에서 이동하는 객체 추적 시 영상 데이터의 전송량을 감소시키는 신경망 계산 시간의 단축 알고리즘을 제안한다. 객체 검출은 디지털화 연속된 영상으로부터 객체 존재 유무를 판단하고, 객체가 존재할 경우 영상 내 객체의 위치, 방향, 크기 등을 알아내는 기술로 정의된다. 그러나 영상 내의 객체는 위치, 크기, 빛의 방향 및 밝기, 장애물 등의 환경적 변화로 인해 객체 모양이 다양해지므로 정확하고 빠른 검출이 어렵다. 따라서 본 논문에서는 신경망을 사용하여 몇 가지 환경적 조건을 극복한 정확하고 빠른 객체 검출 방법을 제안한다. 검색 영역의 축소는 영상 내 색상 영역의 분할과 차영상을 이용하였고, 주성분 분석을 통해 신경망의 입력 벡터를 축소시킴으로써 신경망 수행 시간과 학습 시간을 단축시켰다. 실시간으로 입력되는 동영상에서 모두 실험하였으며, 색상 영역의 분할을 사용할 경우 입력 영상의 칼라 설정의 유무에 따른 검출 성공률의 차를 보였다. 실험 결과에서 보면 제안하는 방법으로써 객체의 움직임을 탐지하였을 때 기존의 방법보다 30% 정도 더 높은 인식 성능을 보여준다.

This paper proposes algorithm to reduce the computing time in a neural network that reduces transmission of data for tracking mobile objects in surveillance networks in terms of detection and communication load. Object Detection can be defined as follows : Given image sequence, which can forom a digitalized image, the goal of object detection is to determine whether or not there is any object in the image, and if present, returns its location, direction, size, and so on. But object in an given image is considerably difficult because location, size, light conditions, obstacle and so on change the overall appearance of objects, thereby making it difficult to detect them rapidly and exactly. Therefore, this paper proposes fast and exact object detection which overcomes some restrictions by using neural network. Proposed system can be object detection irrelevant to obstacle, background and pose rapidly. And neural network calculation time is decreased by reducing input vector size of neural network. Principle Component Analysis can reduce the dimension of data. In the video input in real time from a CCTV was experimented and in case of color segment, the result shows different success rate depending on camera settings. Experimental results show proposed method attains 30% higher recognition performance than the conventional method.

지역 패턴을 정확하게 부호화 하는 방법은 텍스처 분류 연구에 매우 중요한 요소다. 하지만 기존 널리 연구된 LBP기반 방 법들은 잡음에 취약한 근본적인 문제점이 있다. 최근 표정인식 분야에서 에지반응 값과 방향 정보를 활용한 LDP방법이 제안 되었다. LDP방법은 LBP보다 잡음에 강하고 더 많은 정보를 코드에 수용할 수 있는 장점이 있지만 텍스처 분류에 적용하기에 는 치명적인 회전 변화에 민감한 단점이 있다. 본 논문에서는 LDP 방법에 회전 불변 특성을 결합하고 기존 LDP가 가지고 있 던 부호 정보를 수용하지 않은 단점과 밝기 값 차이가 적은 영역에서 의미 없는 코드가 생성되는 단점을 극복한 새로운 지역 패턴 부호화 방법인 Rotation Invariant Local Directional Pattern 방법을 제안한다. 본 논문에서 제안된 방법의 텍스처 분류 성능을 입증하기 위해 널리 사용되는 UIUC, CUReT 데이터 셋에서 텍스처 분류를 수행했다. 그 결과 제안된 RILDP방법이 기존 방법보다 우수한 성능을 보여주었다.

Accurate encoding of local patterns is a very important factor in texture classification. However, LBP based methods w idely studied have fundamental problems that are vulnerable to noise. Recently, LDP method using edge response and dire ction information was proposed in facial expression recognition. LDP is more robust to noise than LBP and can accommod ate more information in it's pattern code, but it has drawbacks that it is sensitive to rotation transforms that are critical to texture classification. In this paper, we propose a new local pattern coding method called Rotation Invariant Local Direc tional Pattern, which combines rotation-invariant transform to LDP. To prove the texture classification performance of the proposed method in this paper, texture classification was performed on the widely used UIUC and CUReT datasets. As a result, the proposed RILDP method showed better performance than the existing methods.

본 논문에서는 질병관리청이 제공하는 퇴원 손상 심층 조사 원자료를 이용하여 환자 발생 패턴을 추출하는 방법 을 제안하고, 이를 기반으로 실험 분석 결과를 제시한다. 제안된 방법은 특정 기간을 중심으로 환자 발생 횟수를 가중치 기반으로 추정하여 발생 패턴을 처리하는 것으로 시작한다. 이후, 시간 시계열 데이터를 빠른 푸리에 변환(Fast Fourier Transform, FFT)을 통해 빈도 성분으로 변환하여 복잡한 시계열 데이터를 명확한 패턴으로 변환한다. 이렇게 변환된 빈도 성분 데이터는 -최근접 이웃(K-Nearest Neighbors, KNN) 알고리즘을 적용하여 여러 발생 패턴으로 분류한다. 본 연구에서는 총 네 가지 주요 패턴으로 분류하였으며, 각 패턴의 특성과 실험 결과를 시각화하여 상세히 제시하였다. 본 연구의 결과는 환자 발생 패턴을 분석하고 이해하는 데 유용한 정보를 제공하며, 관련 데이터 활용 가능성을 높이는 데 기여할 것으로 기대된다.

This paper proposes a method for extracting patient incidence patterns using raw data from in-depth discharge injury investigations provided by the Korea Centers for Disease Control and Prevention and presents the results of experimental analysis based on this method. The proposed approach begins by estimating the frequency of patient incidences using a weight-based method centered around specific periods, thereby processing occurrence patterns. Subsequently, time-series data are transformed into frequency components using Fast Fourier Transform (FFT), enabling the conversion of complex time-series data into distinct patterns. The transformed frequency component data are then classified into multiple incidence patterns using the k-Nearest Neighbors (KNN) algorithm. In this study, a total of four major patterns were identified, and the characteristics of each pattern, along with experimental results, are presented in detail through visualization. The findings of this study provide valuable insights for analyzing and understanding patient incidence patterns and contribute to enhancing the utility of related datasets.

Twin support vector machine (TWSVM) was initially designed for binary classification. However, real-world problems often require the discrimination more than two categories. To tackle multi-class classification problem, in this paper, a multiple least squares twin support vector machine is proposed. Our Multi-LSTSVM solves K quadratic programming problems (QPPs) to obtain K hyperplanes, each problem is similar to binary LSTSVM. Comparison against the Multi-LSSVM, Multi-GEPSVM, Multi-TWSVM and our Multi-LSTSVM on both UCI datasets and ORL, YALE face datasets illustrate the effectiveness of the proposed method.

Pattern classification / recognition is considered as one of the frequently adopted methodology in complex system fault diagnosis. This paper studies a group of methods which can be collectively categorized as signal feature matching based fault diagnosis methodology. This proposed methodology employs dynamic programming to quantify the distance between signals which are processed by morphological filtering. The proposed methods are applied to deterministic fault classification problem in Tennessee Eastman (TE) process to demonstrate the their validity and advantages.

Bed Activity Monitoring System (BAMS) monitors and assess the mobility of people on a bed. This is a useful and critical application for patients with mobility issues after stroke or traumatic brain injury. The system is based on processing of data collected from a piezoelectric pressure sensor for discriminating mobility patterns. There are four different types of motion that were simulated by non-patient volunteers and data was collected. In this paper, two methods were used to extract feature parameters (autoregressive and cepstral coefficients) from the acquired data. Two classification algorithms, Euclidean Distance Measure (EDM) and Weighted Distance Measure (WDM) were used to classify and discriminate the mobility patterns of normal person (healthy subject) from people with mobility issues (patient subjects). Experimental result shows that the recognition rate using cepstral parameters was more effective compare to autoregressive parameters.

Selecting appropriate feature extraction method is absolutely one of the most important factors to archive high classification performance in pattern recognition systems. Among different feature extraction methods proposed for pattern recognition, statistical moments seem to be so promising. Whereas theoretical comparison of the moments is too complicated, in this paper, an experimental evaluation on four well known statistical moments namely Hu invariant moments, Affine invariant moments, Zernike moments, and Pseudo-Zernike moments is presented. Set of different experiments on a binary images dataset consisting of regular, translated, rotated, and scaled Persian printed numerical characters using a nearest neighbor rule classifier has been done and variety of interesting results have been presented. Finally, the results show that Pseudo-Zernike moments outperform the other introduced moments.

DNA computing is a computing paradigm as information storage media and information processing operators. Especially, information processing of DNA computing has emerged for gene expression pattern analysis and disease diagnosis. Recently, we developed a theoretical model for pattern classification, biomolecular perceptron. In the model, each pattern is a solution of various nucleotide strands. The patterns of the DNA species are summed with corresponding weight factors, where the result is used for the pattern’s classification. The calculation of this weighted summation is performed via competitive hybridization reactions between input molecules and differentially labeled probe molecules (molecular beacon), and the calculation is detected as fluorescence signals. Here, we demonstrate the experimental feasibility of pattern classification. We perform disease type discrimination using the subsequence of the acute leukemia marker genes. The results show the possibility of our method as a pattern classification tool and diagnosis tool using gene expression. Because of the absence of information conversion steps, diagnosis process can be simplified.

This paper proposes a Multiclass Least Squares Twin Support Vector Machine (MLSTSVM) classifier for multi-class classification problems. The formulation of MLSTSVM is obtained by extending the formulation of recently proposed binary Least Squares Twin Support Vector Machine (LSTSVM) classifier. For M-class classification problem, the proposed classifier seeks M-non parallel hyper-planes, one for each class, by solving M-linear equations. A regularization term is also added to improve the generalization ability. MLSTSVM works well for both linear and non-linear type of datasets. It is relatively simple and fast algorithm as compared to the other existing approaches. The performance of proposed approach has been evaluated on twelve benchmark datasets. The experimental result demonstrates the validity of proposed MLSTSVM classifier as compared to the typical multi-classifiers based on ‘Support Vector Machine’ and ‘Twin Support Vector Machine’. Statistical analysis of the proposed classifier with existing classifiers is also performed by using Friedman’s Test statistic and Nemenyi post hoc techniques.

We introduce a new category of fuzzy neural networks with multiple-output based on fuzzy clustering algorithm, especially, fuzzy c-means clustering algorithm (FCM-based FNNm) for pattern classification in this paper. The premise part of the rules of the proposed networks is realized with the aid of the scatter partition of input space generated by FCM clustering algorithm. The partitioned local spaces describe the fuzzy rules and the number of the partitioned local spaces is equal to the number of clusters. Due to these characteristics, we may alleviate the problem of the curse of dimensionality. The consequence part of the rules is represented by polynomial functions with multiple-output for pattern classification. And the coefficients of the polynomial functions are learned by back propagation algorithm. To optimize the parameters of the proposed FCM-based FNNm we consider real-coded genetic algorithms. The proposed networks are evaluated with the use of numerical experimentation.

A plethora of big data applications are emerging and being researched in the computer science community which require online classification and pattern recognition of huge data pools collected from sensor networks, image and video systems, online forum platforms, medical agencies etc. However, as an NP hard issue data mining techniques are facing with lots of difficulties. To deal with the hardship, we conduct research on the novel algorithm for data mining and knowledge discovery through network entropy. We firstly introduce necessary data analysis techniques such as support vector machine, neural network and decision tree methods. Later, we analyze the organizational structure of network graphical pattern with the knowledge of machine learning methodology and graph theory. Eventually, our modified method is finalized with decision and validation implementation. The simulation results of our approach on different databases show the feasibility and effectiveness of our proposed framework. As the final part, we provide our conclusion and prospect.

RFID technology has been widely used in many fields. Lots of information could be obtained from the enormous RFID devices. Therefore, classification of the trajectory of these moving objects is significant to predict the trends of the moving objects. This paper introduces an innovative algorithm, using pattern recognition approach. This algorithm could be divided into several steps. First of all, the coarse-fine layer classification approach is used for clustering the trajectories. That aims to cut down the trajectories according to the difference phases. Secondly, in the sub-set of the trajectory, searching set is established through detecting the neighbor domains which have been classified at different phases. Finally, the hierarchical classification approach is used for classifying the sub-trajectory. By using the approach, the experimental results imply the feasibility and practicality of the proposed approach on figuring out the familiar trajectory of the RFID-enabled moving objects. It is observed that With the increasing of neighbor value , the linear trends is obvious from the figure, thus, the neighbor query decreases and the magnification ratio outperforms to TRACLUS method. Additionally, the proposed algorithm uses coarse-fine strategy at phases-to-phases, saving the time spend on large number of distance calculation at various phase.

In this work, we have presented a hierarchical scheme for classifying music data. Instead of dealing with large variety of features, proposed scheme relies on MFCC and its variants which are introduced at the different stages to satisfy the need. At the top level music is classified as song (music with voice) and instrumental (music without voice) based on MFCC. Subsequently, instrumental signals and songs are classified based on instrument type and genres respectively. Hierarchical approach has been followed for such detailed categorization. Using two-stage process, instrumental signals are identified as one of the four types namely, string, woodwind, percussion or keyboard. Wavelet and MFCC based features are used for this purpose. For song classification, at first level signals are categorized as classical or non-classical(popular) ones by capturing the MFCC pattern present in the high sub-band of wavelet decomposed signal. At second level, we consider the task of further classification of popular songs into various genres like Pop, Jazz, Bhangra (an Indian genre) based on amplitude variation pattern. RANSAC has been utilized as the classifier at all stages. Experimental result indicates the effectiveness of the proposed schemes.

Study of different patterns on a local neighborhood of a texture plays an important role in characterization, and classification of the textures. The present paper proposes a method for measuring the occurrence factor of patterns on a randomly thresholded binary image. For this process eight simple patterns are chosen on a 3×3 neighborhood. The simple patterns are chosen in such a way that any complex pattern can be formed by grouping one or more of these simple patterns. The pattern occurrence factor of different binary images is also compared with the actual binary texture image. The experimental results on sixty four textures indicate good comparison of variation of occurrence in these patterns on different binary images of random threshold

청소년들은 ‘위기’를 겪게 되는데, 이 위기를 어떻게 보내느냐에 따라 개개인의 일생에 큰 영향을 미치게 된다. 현대사회는 청소년의 발달과 적응에 장애가 될 수 있는 다양한 위기상황이 존재하므로 기독교 교육적 차원에서 위기의 청소년을 이해하고 도울 수 있는 방안을 모색할 때이다. 선행연구를 통하여 위기청소년의 개념을 정리하였고 그들의 특징 과 문제점을 설명하였다. 현재 학교현장과 교육기관에서 사용되어지는 시스템의 분석을 통해 청소년의 위기유형을 분류하는 기준을 분석하였으며 기독교교육을 통한 가정사역적 접근을 통해 교회에서 품어야 할 위기청소년의 다양한 지원방법을 제시하고자 하였다. 본 연구에서 학생의 위기유형은 크게 세 가지 유형(가정위기, 감정ㆍ정서위기, 학교위 기)로 분류된다. 가정위기는 정서대화, 경제문제, 가정폭력, 결손 및 조손가정의 문제들을 포함하고 감정ㆍ정서위기는 우울, 불안, 충동조절, 스트레스등의 문제들을 포함한다. 이러 한 문제들은 학교위기로 나타나는데, 학생의 개인적인 위기유형으로 우울, 불안, 자해, 자 살등의 정신건강적인 양상들과 인터넷 과몰입, 음주ㆍ흡연의 모습을 보이며, 학생의 사회 적인 문제로는 집단 따돌림, 도벽, 금품갈취, 학교폭력, 학교부적응, 품행장애, ADHD, 성폭 력 등의 사회적 문제들이 야기될 수 있다고 하였다. 청소년의 3가지 위기유형을 극복하기 위하여 긍정적 개인요인을 통한 위기지원, 가정사역에서 가정을 통한 위기지원, 교회 및 사회기관을 통한 위기지원, 전문상담치료를 통한 위기지원으로 제시하였다.

Christian Educational Approach for Crisis Pattern Classification in School Counseling: Family Life Education Mila Park Adjunct Professor, Sungkyul University Every youth has gone through the crisis period, and how to undergo this time has made big influence to their entire life. Modern society has offered various crisis environment to obstacle the youths’ development and adaptation. This study tried to understand and help the youths who faced crisis situations with Christian educational approach. The researcher proposed the concepts of youth crises and their characteristics and problems through the literature reviews. This research classified the students’ crisis patterns with three categories: family crisis, emotional crisis, and school crisis. Family crisis included communication problems, financial problems, violences, parentless family, and single family problems. Emotional crisis included depression, anxiety, impulse control disorder, and stress. Family and emotional crisis problems are turned up to the school crisis problems. Individual student’s crisis patterns appeared in psychiatric and health problems such as depression, anxiety, self-inflicted wound, suicide, Internet addiction, and alcoholic and smoking problems. These problems are connected to the societal problems such as ostracism, kleptomaniac, extortion, violence, school maladjustment, ADHD, sexual violence. 4. 학교상담현장에서의 위기유형분류와 기독교 가정교육(사역)적 접근 107 This research suggested various supporting alternatives to overcome youth students’ three crisis patterns. Affirmative self-esteem, family life educational supporting, local church and social institution supporting, and specialized counseling treatment are proposed to cure the youths’ family, emotional, and school crisis problems.

데이터 마이닝의 문서분류 기술에서 발전된 오피니언 마이닝은 이제 국외뿐만 아니라 국내 산업에서 중요한 관심분야로 자리잡아가고 있다. 오피니언 마이닝의 핵심은 문서에서 감정 단어를 추출하여 긍정/부정 여부를 얼마나 정확하게 판별하느냐를 평가하는 것이다. 국내에서도 이에 관련된 많은 연구가 이루어 졌으나 아직 실용적으로 적용 할 만큼의 분류 정확도를 보이지 않고 있다. 한국어의 경우 비문법적 표현, 감정단어의 다양성 등으로 인해 문서의 극성을 판별하기가 쉽지 않기 때문이다. 본 논문에서는 문법적 요소를 최대한 배제하고 단어패턴의 빈도만을 고려한 새로운 오피니언 문서 분류기법을 제안한다. 제안된 방법에서는 문서를 단어들의 리스트로 추상화한 후, 패턴들의 빈 도를 이용하여 기계학습 알고리즘을 적용한다. 이후에 적절한 스코어 함수를 적용하여 문서의 극성을 판별한다. 또한 제안된 기법의 정확도를 평가하기 위해서 실험결과를 제시한다.

An opinion mining technique which was developed from document classification in area of data mining now becomes a common interest in domestic as well as international industries. The core of opinion mining is to decide precisely whether an opinion document is a positive or negative one. Although many related approaches have been previously proposed, a classification accuracy was not satisfiable enough to applying them in practical applications. A opinion documents written in Korean are not easy to determine a polarity automatically because they often include various and ungrammatical words in expressing subjective opinions. Proposed in this paper is a new approach of classification of opinion documents, which considers only a frequency of word patterns and excludes the grammatical factors as much as possible. In proposed method, we express a document into a bag of words and then apply a learning algorithm using a frequency of word patterns, and finally decide the polarity of the document using a score function. Additionally, we also present the experiment results for evaluating the accuracy of the proposed method.

필적 감정서에서 특이성과 빈도의 가중치가 높은 글자로 선택되어 사용된 글자를 자음별로 구분하였을 때 ‘ㅅ, ㅈ 및 ㅎ’의 자음에서 추출된 세부 특징의 수도 많고 유형별로 구분되는 수도 다양하여 감정에 많이 활용되는 것을 확인하였다. 필적 감정서에 특이성과 빈도의 가중치가 높아 세부특징 글자에 사용된 글자는 1건에서만 나타나는 글자가 82% 정도로 특정한 글자만 가지고 감정을 하는 것은 적합하지 않음을 확인하였다. 한글 자음을 유형별로 1차 및 2차로 세부 분류하고 가중치를 결정한 결과, 자음 ‘ㄷ, ㄹ, ㅂ 및 ㅌ’의 유형만 가중치가 2로 결정되었다. 그리고 나머지는 모두 가중치가 3으로 결정되었으며 추출된 글자 자체가 특이성과 빈도수가 높은 글자를 감정서에 활용한 것이므로 이 같은 결과가 나타난 것을 알 수 있었다.

As a lot of handwriting characteristics have been extracted from consonants “ㅅ, ㅈ, and ㅎ” and these consonants can be divided into many types, the weight score of the frequency and uniqueness for the consonants “ㅅ, ㅈ, and ㅎ” have been high, resulting in frequent use of the consonants for handwriting examination. The percentage of a certain letter with high uniqueness and frequency that appeared during a single handwriting examination is 82%, implying that conducting examinations using only 1 letter is unwise. Korean consonants were classified into two groups by the consonant's type and then weight score of each consonant was analyzed. The consonants “ㄷ, ㄹ, ㅂ, and ㅌ” received a weight score of 2 and the rest received a weight score of 3. The analyzed characters themselves, which have high uniqueness and frequency, were utilized in the examination report, resulting in such results.

패턴 선택은 주어진 전체 학습 패턴으로부터 일부 유용한 패턴들을 추출하여 학습에 이용하는 방법이다. 분류 문제에서 유용한 학습패턴은 클래스들 간의 분류경계에 근접하거나 클래스의 중심 부근에 속한 패턴을 정상패턴이라고 한다. 이러한 패턴 중에서 노이즈 패턴이 포함된다면 분류 모델의 성능을 저하시킨다. 제안하는 방법은 노이즈를 제거하는 일종의 전처리 작업으로서 기존의 샘플링 방법을 사용하지 않고, 비감독 학습을 도입하여 노이즈 데이터를 제거한다. 그 결과 선택된 데이터로 학습할 경우, 학습 데이터 수 감소로 인한 계산 시간 단축 및 메모리 공간 절약의 효과 그리고 클래스 불균형 문제에서 불균형 비율을 감소시키는 효과가 있다. 제안하는 방법의 효율성을 알아보기 위하여 인공 데이터와 실제 문제에 적용하였다. 적용 결과 breast cancer 데이터는 전체 데이터수의 54%를 제거하였어도 정확률, 재현율, F1 값이 모두 향상되었다.

Pattern selection algorithms are to select some useful patterns from all learning patterns and use them as training patters. Useful pattern is a pattern that contributes much to the learning. For classification problems, useful patters are those near class boundary or class center. Noise patterns will decrease the performance of classification model. In this paper, we propose a pre-processing technique to eliminate those noise patterns. This is based on un-supervised learning technique, which is different from traditional sampling techniques. We will have several benefits when model is learned from useful patterns only: decrease in both computation time and space requirements while learning as well as the decrease of class imbalance. To show the effectiveness of proposed techniques, we applied it to one synthetic data set and two real-world benchmark data sets. For the breast cancer case, we achieved the increase of all of precision, recall, and F1 values even if data set is reduced to less than 50%.