• International Journal of Internet, Broadcasting and Communication
• /
• 제11권4호
• /
• pp.37-42
• /
• 2019

In this paper, we explore the details of three classic data augmentation methods and two generative model based oversampling methods. The three classic data augmentation methods are random sampling (RANDOM), Synthetic Minority Over-sampling Technique (SMOTE), and Adaptive Synthetic Sampling (ADASYN). The two generative model based oversampling methods are Conditional Generative Adversarial Network (CGAN) and Wasserstein Generative Adversarial Network (WGAN). In imbalanced data, the whole instances are divided into majority class and minority class, where majority class occupies most of the instances in the training set and minority class only includes a few instances. Generative models have their own advantages when they are used to generate more plausible samples referring to the distribution of the minority class. We also adopt CGAN to compare the data augmentation performance with other methods. The experimental results show that WGAN-based oversampling technique is more stable than other approaches (RANDOM, SMOTE, ADASYN and CGAN) even with the very limited training datasets. However, when the imbalanced ratio is too small, generative model based approaches cannot achieve satisfying performance than the conventional data augmentation techniques. These results suggest us one of future research directions.

• 지능정보연구
• /
• 제21권2호
• /
• pp.173-190
• /
• 2015

부도는 막대한 사회적, 경제적 손실을 야기할 수 있으므로, 미리 부도여부를 정확하게 예측하여 선제 대응하는 것은 경영분야에서 대단히 중요한 의사결정문제 중 하나이다. 이에 지능정보시스템 분야에서도 그간 기업의 재무 데이터에 기반해 부도예측을 개선하기 위한 노력을 기울여왔는데, 안타깝게도 기존의 연구들은 대부분 분류모형의 성능 개선을 통해 예측 정확도를 개선하는 것에만 주로 초점을 맞추어 다른 요소들을 충분히 고려하지 못했다는 한계가 있다. 이러한 배경에서 본 연구는 부도예측 모형의 정확도를 개선하기 위한 방편으로 새로운 데이터 전처리 방법, 그 중에서도 효과적인 표본추출 방법을 제안하고자 한다. 일반적으로 부도예측을 위해 사용되는 데이터들은 극심한 데이터 불균형 문제에 노출되어 있는데, 본 연구에서는 k-reverse nearest neighbor(k-RNN)와 one-class support vector machine(OCSVM) 방법을 결합한 하이브리드 언더샘플링(hybrid under-sampling) 접근법을 통해 이같은 데이터 불균형 문제를 해결하고자 하였다. 본 연구에서 제안한 접근법에서 k-RNN은 이상치를 효과적으로 제거할 수 있으며, OCSVM은 다수를 구성하는 등급의 데이터로부터 정보량이 풍부한 표본만 효과적으로 선택할 수 있는 수단으로 활용될 수 있다. 제안된 기법의 성능을 검증하기 위해, 본 연구에서는 국내 한 은행의 비외감기업 부도예측모형 구축에 제안 기법을 적용해 본 뒤, 일반적으로 많이 사용되는 랜덤샘플링(random sampling)과 제안 기법의 성능을 비교해 보았다. 그 결과, 로지스틱 회귀분석, 판별분석, 의사결정나무, SVM 등 대다수의 분류모형에 있어 분류 정확도가 개선됨을 확인할 수 있었으며, 모든 분류모형에 있어 부정 오류, 즉 부실기업을 정상으로 예측하는 오류율이 크게 감소함을 확인할 수 있었다.

• Nuclear Engineering and Technology
• /
• 제51권8호
• /
• pp.1951-1956
• /
• 2019

As a part of the preliminary decommissioning plan of KEPCO-NF fuel fabrication facility, DCGLs of three target radionuclides, ²³⁴U, ²³⁵U, and ²³⁸U, were derived using RESRAD-BUILD code and contaminated areas of the facility were classified based on contamination levels from the derived DCGLs. From code simulations, one-room modeling results showed that the grinding room in building #2 was the most restrictive (DCGL_gross = 10493.01 Bq/㎡). The DCGL_gross results in contaminated areas from one-room modeling were slightly more conservative than three-room modeling. Prior to the code simulation, field survey and measurements conducted by each survey unit. For a conservative approach, the most restrictive DCGL_gross in each survey unit was taken as a reference to classify the contaminated areas of the facility. Accordingly, seven rooms and 37 rooms in the nuclear-fuel buildings were classified as Class 1 and Class 2, respectively. As expected, fuel material handling and processing rooms such as the grinding room, sintering room, compressing room, and powder collecting room were included in the Class 1 area.

• 전기전자학회논문지
• /
• 제25권2호
• /
• pp.273-279
• /
• 2021

본 논문에서는 딥러닝 구조를 분할을 이용한 개별 학습을 수행하여 학습시간을 단축하는 알고리즘을 제안한다. 제안하는 알고리즘은 망 분류 기점 설정 과정, 특징 벡터 추출 과정, 특징 노이즈 제거 과정, 클래스 분류 과정 등의 4가지 과정으로 구성된다. 첫 번째로 망 분류 기점 설정 과정에서는 효과적인 특징 벡터 추출을 위한 망 구조의 분할 기점을 설정한다. 두 번째로 특징 벡터 추출 과정에서는 기존에 학습한 가중치를 사용하여 추가 학습 없이 특징 벡터를 추출한다. 세 번째로 특징 노이즈 제거 과정에서는 추출된 특징 벡터를 입력받아 각 클래스의 출력값을 학습하여 데이터의 노이즈를 제거한다. 네 번째로 클래스 분류 과정에서는 노이즈가 제거된 특징 벡터를 입력받아 다층 퍼셉트론 구조에 입력하고 이를 출력하고 학습한다. 제안된 알고리즘의 성능을 평가하기 위하여 Extended Yale B 얼굴 데이터베이스를 사용하여 실험 하였다. 실험 결과, 1회 학습에 소요되는 시간의 경우 제안하는 알고리즘이 기존 알고리즘 기준 40.7% 단축하였다. 또한 목표 인식률까지 학습 횟수가 기존 알고리즘과 비교하여 단축하였다. 실험결과를 통해 1회 학습시간과 전체 학습시간을 감소시켜 기존의 알고리즘보다 향상됨을 확인하였다.

• Journal of Oral Medicine and Pain
• /
• 제24권1호
• /
• pp.37-47
• /
• 1999

This study was performed to investigate the factors related to vibration of temporomandibular joint during mandibular opening movement. For this study, 144 patients with temporomandibular disorders were randomly selected. Angle's classification, lateral guidance pattern, range of maximal mouth opening, preferred chewing side, and affected side were investigated clinically. Mandibular torque rotational movement during opening was recorded with $BioEGN^{(R)}$ and vibration of temporomandibular joint during opening was recorded with $Sonopak^{(R)}$. After clinical diagnosis was made, visual analogue scale(VAS) was used for evaluation of clinical progress of the subject's chief complaints. The author calculated VAS treatment index(VAS Ti) from the record of VAS. The more VAS Ti was, the less remission of subjective symptom was, The data were analyzed with SAS/Stat program and the results of this study were as follows: 1. There were no significant difference in all the variables of joint vibration by age and sex. 2. Integral and peak amplitude in patients of Angle's class I were higher than those of class II or III patients. Integral in patients of group function was higher than that in patients of canine guidance or other types of lateral excursion. 3. As to Angle's classification or lateral guidance type, there were almost not significant difference between subgroup of same class or type and subgroup of different class or type on both sides. And there were also almost not difference between one side and the other side related to preferred chewing side or affected side. 4. Patients with disk displacement with reduction showed higher value of integral and peak amplitude than any other patients. 5. Joint vibration variables significantly correlated with VAS Ti of pain. with clinical range of mouth opening, and with ingredients of mandibular torque rotational movement.

• 한국의상디자인학회지
• /
• 제7권2호
• /
• pp.75-85
• /
• 2005

This study analyzed the factors of change of women's underwear in the Renaissance, which had a sudden change of shape. First, the spirit of the Renaissance was focused on human-attached importance to glamorous beauty of the body and pursued the glamorous well-proportioned figure as the ideal of the human body. This expressed a woman's beautiful curved lines by reduction of their waist size and emphasis of breast and hip lines. It also created and emphasized one's physical figure, which is a characteristic of sex. Also the materials and size of underwear cleared up the classification of class. Second, through the development of weaving techniques, more textiles were produced, from linen, the most common material used, to silk in underwear and stomacher, and chemise, which was made more splendid be devising elaborate embroidery techniques. Third, as we know that the farthingale was devised in Spain and transmitted to France because of the prevalence of printing and trade, where it changed and developed to more convenient style, this shows that information interchange was active, and we can see the phenomenon that it is developing continually through the prevalence of printing and trade.

• 한국대기환경학회지
• /
• 제13권1호
• /
• pp.51-63
• /
• 1997

Total suspended particulate matters in the ambient air were analyzed for eight chemical elements (Ca, Co, Cu, Fe, Mn, Pb, Si, and Zn) using an x-ray fluorescence spectrometry (XRF) at the Kyung Hee University - Suwon Campus during 1989 to 1994. To use these data as basis for source identification study, membership of each sample was selected to represent one of the well defined sample groups. The data sets consisting of 83 objects and 8 variables were initially separated into two groups, fine (d$_{p}$<3.3 ${\mu}{\textrm}{m}$) and coarse particle groups (d$_{p}$>3.3 ${\mu}{\textrm}{m}$). A hierarchical clustering method was examined to obtain possible member of homogeneous sample classes for each of the two groups by transforming raw data and by applying various distances. A disjoint principal component analysis was then used to define homogeneous sample classes after deleting outliers. Each of five homogeneous sample classes was determined for the fine and the coarse particle group, respectively. The data were properly classified via an application of logarithmic transformation and Euclidean distance concept. After determining homogeneous classes, correlation coefficients among eight chemical variables within all the homogeneous classes for calculated and meteorological variables (temperature. relative humidity, wind speed, wind direction, and precipitation) were examined as well to intensively interpret environmental factors influencing the characteristics of each class for each group. According to our analysis, we found that each class had its own distinct seasonal pattern that was affected most sensitively by wind direction.ion.

• 한국콘텐츠학회논문지
• /
• 제9권7호
• /
• pp.141-148
• /
• 2009

최근에 클래스 분포의 불균형이 심한 데이터의 학습 문제가 그 중요도에 비하여 만족할만한 성능을 얻기 어려운 관계로 관심이 고조되고 있다. 이 문제에 대한 접근 방법은 데이터 레벨의 불균형 해소, 알고리즘 레벨에서의 비용함수 도입, 인식기의 앙상블에 의한 성능향상 등으로 분류된다. 이 논문은 알고리즘 레벨의 접근 방법으로써, 다층퍼셉트론 신경회로망에 고차의 오차함수를 사용하여 불균형 데이터를 학습하는 방법을 제시한다. 즉, 소수클래스의 학습을 강화시키고 다수 클래스의 학습을 약화시키는 형태로 가 중치를 변경시킨다. 클래스 불균형이 심한 유방암 검사와 갑상선 진단 데이터의 학습을 통하여 제안한 방법이 MSE(mean-squaerd error), 2단계 방법 및 문턱조정 방법보다 우수함을 확인한다.

• 한국정보처리학회논문지
• /
• 제5권1호
• /
• pp.103-110
• /
• 1998

기존의 규칙베이스 추론(Rule-Based REasoning : RBR)과 사례베이스 추론 (Case-Base : CB)가 통합되어 추론되고 있지만, 많은 수의 규칙(Rule)과 사례(Case)에 의해 추론 시간이 많이 걸리는 단점이 있다. 본 논문에서는 이런 단점을 해결하기 위하여, 다중 의미 또는 불확실한 지식을 쉽게 표현할 수 있는 라프집합 (Rough Set)을 이용하여 RB와 CB를 간략화한 새로운 추론 방법을 제안한다. 라프집합의 식별(classification)과 근사(aprroximation)개념을 이용하여, RB와 CB를 통치 클래스(equivalence class)로 분류하여 각각을 각략화하고, 간략화된 RB와 CB를 이용하여 통합 추론하여, 상호 보완적인 역할에 의해 결정 해를 얻고자 하는 것이다.

• Journal of Chest Surgery
• /
• 제25권10호
• /
• pp.1055-1060
• /
• 1992

Clinical experience of 21 patients with infective endocarditis was reviewed. Endocarditis involved the left-sided valve in 16 cases, the right-sided valve in 2, and PDA in the remaining 3 patients. Valve abnormalities included leaflet perforation in 9 patients, chordal rupture in 2,; annular abscess in 6; and aorticoleft atnal perforation in 2. Sixteen patients underwent valve replacement[aortic valve replacement in 7 patients, mitral replacement in 4 and double valve replacement in 5], two had VSD closure with pulmonary valve excision, three had ductus arteriousus closure. The patients were classified into two groups. I ] Healed endocarditis group: including the patients who had completed a planned cou-rseof antibiotic therapy[N=10], II ] Active endocarditis group: patients in which operations were performed prior to completetion of antibiotic treatment course[N=11]. The indications for operation included congestive heart failure, embolism, and persistent sepsis. Organisms were predominantly streptococcus[N=5] and staphylococcus [N=4] followed by candida, moraxella, and E-coli. By NYHA functional classification, all patients were in Class III or IV preoperatively. There was only one operative mortality in patient from group II. All patients substantially, improved postoperatively with NYHA classification in class I or II. This study shows that early surgical intervention in patients with active endocarditis has desirable outcome.