• International Journal of Fuzzy Logic and Intelligent Systems
• /
• 제5권3호
• /
• pp.263-267
• /
• 2005

Building classifiers for financial real-world classification problems is often plagued by severely overlapping and highly skewed class distribution. New performance measures such as receiver operating characteristic (ROC) curve and area under ROC curve (AUC) have been recently introduced in evaluating and building classifiers for those kind of problems. They are, however, in-effective to evaluation of classifier's discrimination performance in a particular class of the classification problems that interests lie in only a local operating range of the classifier, In this paper, a new method is proposed that enables us to directly improve classifier's discrimination performance at a desired local operating range by defining and optimizing a partial area under ROC curve or domain-specific curve, which is difficult to achieve with conventional classification accuracy based learning methods. The effectiveness of the proposed approach is demonstrated in terms of fraud detection capability in a real-world fraud detection problem compared with the MSE-based approach.

• Communications for Statistical Applications and Methods
• /
• 제4권1호
• /
• pp.229-241
• /
• 1997

We consider discrimination curve and minimum dwell time for Poisson distribution and Poisson-power function distribution. Let the random variable X has Poisson distribution with mean .lambda.. For the hypothesis testing H$\_$0/:.lambda. = t vs. H$\_$1/:.lambda. = d (d$\_$0/ if X.leq.c. Since a critical value c can not be determined to satisfy both types of errors .alpha. and .beta., we considered discrimination curve that gives the maximum d such that it can be discriminated from t for a given .alpha. and .beta.. We also considered an algorithm to compute the minimum dwell time which is needed to discriminate at the given .alpha. and .beta. for the Poisson counts and proved its convergence property. For the Poisson-power function distribution, we reject H$\_$0/ if X.leq..'{c}.. Since a critical value .'{c}. can not be determined to satisfy both .alpha. and .beta., similar to the Poisson case we considered discrimination curve and computation algorithm to find the minimum dwell time for the Poisson-power function distribution. We prosent this algorithm and an example of computation. It is found that the minimum dwell time algorithm fails for the Poisson-power function distribution if the aiming error variance .sigma.$\^$2/$\_$2/ is too large relative to the variance .sigma.$\^$2/$\_$1/ of the Gaussian distribution of intensity. In other words, if .ell. is too small, we can not find the minimum dwell time for a given .alpha. and .beta..

• Communications for Statistical Applications and Methods
• /
• 제20권3호
• /
• pp.169-174
• /
• 2013

The ROC curve is drawn with two conditional cumulative distribution functions (or survival functions) of the univariate random variable. In this work, we consider joint cumulative distribution functions of k random variables, and suggest a ROC curve for multivariate random variables. With regard to the values on the line, which passes through two mean vectors of dichotomous states, a joint cumulative distribution function can be regarded as a function of the univariate variable. After this function is modified to satisfy the properties of the cumulative distribution function, a ROC curve might be derived; moreover, some illustrative examples are demonstrated.

• 전자공학회논문지CI
• /
• 제45권6호
• /
• pp.179-186
• /
• 2008

본 논문에서는 베이지안 알고리즘이 불균형 데이터의 학습 시 나타나는 현상을 분석하고 성능 평가 방법을 비교하였다. 사전 데이터 분포를 가정하고 불균형 데이터 비율과 분류 복잡도에 따라 발생된 분류 문제에 대해 베이지안 학습을 수행하였다. 실험 결과는 ROC(Receiver Operator Characteristic)와 PR(Precision-Recall) 평가 방법의 AUC(Area Under the Curve)를 계사하여 불균형 데이터 비율과 분류 복잡도에 따라 분석되었다. 비교 분석에서 불균형 비율은 기 수행된 연구 결과와 같이 베이지안 학습에 영향을 주었으며, 높은 분류 복잡도로부터 나타나는 데이터 중복은 학습 성능을 방해하는 요인으로 확인되었다. PR 평가의 AUC는 높은 분류 복잡도와 높은 불균형 데이터 비율에서 ROC 평가의 AUC보다 학습 성능의 차이가 크게 나타났다. 그러나 낮은 분류 복잡도와 낮은 불균형 데이터 비율의 문제에서 두 측정 방법의 학습 성능의 차이는 미비하거나 비슷하였다. 이러한 결과로부터 PR 평가의 AUC는 클래스 불균형 문제의 학습 모델의 설계와 오분류 비용을 고려한 최적의 학습기를 결정하는데 도움을 줄 수 있다.

• 한국산업정보학회논문지
• /
• 제23권1호
• /
• pp.1-12
• /
• 2018

본 논문은 듀얼 에너지 X-ray 검색기의 영상을 이용한 물질의 추정 방법 알고리즘을 제안한다. 물질 추정 알고리즘으로 많이 사용되는 기존 4가지 분별 곡선 이외에 로그 함수를 사용한 새로운 분별곡선을 이용하여 물질을 분류한다. 여기에 기존의 선형 보간을 이용한 원자번호 추정 방법이 아닌 확률분포를 이용한 원자번호 추정 방법을 제시한다. 확률분포를 이용한 가중치 계산에는 근접한 두 기준물질을 사용하는 방법과 모든 기준물질을 사용하는 방식, 2가지 방식을 실험하였다. 확률분포를 가중치로 사용하여 물질의 원자번호를 추정 할 경우 기존의 방법보다 더 정확한 원자번호 추정 결과를 나타내었다. 추정된 원자번호를 육안으로 확인하기 위하여 HSI 모델을 이용하여 결과영상에 채색하였다.

• 한국생활과학회지
• /
• 제14권3호
• /
• pp.467-473
• /
• 2005

The integrative Internet program was developed for the selection of optimal brassiere components according to the breast types of middle-aged women. In this program, a customer is classified into a certain group through three steps of a discriminant analysis. Three variables used in the analysis include breast volume, radii of curvature of under-breast curve, and a distance between inner breast points. Using individual data of three variables, the optimal brassiere components, i.e. brassiere cup size, curvature of front panel and wire, distance between cups, are suggested for each customer. Discrimination of breast types using only 2D measurements is also included for those who do not have easy access to a 3D measurement device.

• 한국자기공명학회논문지
• /
• 제22권4호
• /
• pp.149-157
• /
• 2018

Diabetes is known to be one of common causes for several types of peripheral nerve damage. Diabetic neuropathy (DN) is a significant complication lowering the quality of life that can be frequently found in diabetes patients. In this study, the metabolomic characteristic of DN and Diabetes was investigated with NMR spectroscopy. The sera samples were collected from DN patients, Diabetes patients, and healthy volunteers. Based on the pair-wise comparison, three metabolites were found to be noticeable: glucose, obviously, was upregulated both in DN patients (DNP) and Diabetes. Citrate is also increased in both diseases. However, the dietary nutrient and biosynthesized metabolite from glucose, ascorbate, was elevated only in DNP, compared to healthy control. The multivariate model of OPLS-DA clearly showed the group separation between healthy control-DNP and healthy control-Diabetes. The most significant metabolites that contributed the group separation included glucose, citrate, ascorbate, and lactate. Lactate did not show the statistical significance of change in t-test while it tends to down-regulated both in DNP and Diabetes. We also conducted the ROC curve analysis to make a multivariate model for discrimination of healthy control and diseases with the identified three metabolites. As a result, the discrimination model between healthy control and DNP (or Diabetes) was successful while the model between DNP and Diabetes was not satisfactory for discrimination. In addition, multiple combinations of lactate and citrate in the OPLS-DA model of healthy control and diabetes group (DNP + Diabetes patients) gave good ROC value of 0.952, which imply these two metabolites could be used for diagnosis of Diabetes without glucose information.

• 한국환경보건학회지
• /
• 제47권4호
• /
• pp.320-329
• /
• 2021

Background: Smoking exposure may be objectively assessed through specific biomarkers. The most common biomarker for smoking is cotinine concentration in urine, and setting an optimal cut-off point can accurately classify smoking status. Such a cut-off point for Korean adolescents has never been studied. Objectives: The aim of this study was to determine a cut-off point for urinary cotinine concentration for the discrimination of smoking in adolescents. Methods: Participants were adolescents aged 13~18 years who participated in the third cycle of the Korean National Environmental Health Survey. We used urine samples to confirm the level of cotinine concentrations. Smoking status was determined by self-reported questionnaire. We identified the optimal cotinine cut-off point for discriminating smoking status using receiver operating characteristic curve analysis. Results: Of the 904 participants, 28 (3.1%) were smokers, among whom 20 (71.4%) were male. The median urinary cotinine concentrations in smokers was 218 ㎍/L (male: 215 ㎍/L, female: 303 ㎍/L), and that in non-smokers was 1.31 ㎍/L (male: 1.46 ㎍/L, female: 1.18 ㎍/L). We found significant differences in urinary cotinine concentration according to smoking status and sex (p<0.001). Urinary cotinine concentrations performed well for identifying smoking adolescents [area under the curve: 0.954 (male: 0.963, female: 0.908)]. The cut-off that optimally distinguished smokers from non-smokers was 39.85 ㎍/L (sensitivity: 89.3%, specificity: 97.4%). Male [39.85 ㎍/L (sensitivity: 90.0%, specificity: 94.9%)] had a different optimal cut-off point than female [26.26 ㎍/L (sensitivity: 87.5%, specificity: 99.6%)]. Conclusions: This study determined a cut-off point for urinary cotinine of 39.85 ㎍/L (male: 39.85 ㎍/L, female: 26.26 ㎍/L) to distinguish smokers from non-smokers in adolescents.

• 자동차안전학회지
• /
• 제4권2호
• /
• pp.22-26
• /
• 2012

Prasad and Mertz published head injury risk curves for skull fracture and for Abbreviated Injury Scale (AIS) ${\geq}4$ brain injury due to forehead impacts based on the 15 ms HIC criterion. KNCAP adopted the HIC36 criterion for the male dummy and the HIC15 criterion for the female dummy. In this paper, it was studied that which of the HIC15 and HIC36 was more effective for the male dummy head injury evaluation. The frontal US-NCAP data for the 7 vehicles from the NHTSA test database were used to evaluate the head injuries. In the case of using the HIC15 and evaluation range 250~700, the discrimination of the rating for the occupant head injury was increased.

• 전자공학회논문지B
• /
• 제31B권1호
• /
• pp.99-106
• /
• 1994

In this paper, we develop an algorithm which recognizes Ilpa-style shorthand characters by on-line. It discriminates the structure of characters using coordinates which are measured by tablet board, then it outputs the recognized characters using the fuzzy inference rules. Shorthand characters have several forms, in which an initial or a middle sound depends on angle and length while a last sound is treated as a hook. We apply fuzzy inference rules to the discrimination of the length, the angle, the curve, and the straight line. We also built up a set of standard character codes in order to reduce the processing time.