Earticle

In this paper, we propose a method to detect concept drift by applying Convolutional Neural Network (CNN) in a data stream environment. Since the conventional method compares only the final output value of the CNN and detects it as a concept drift if there is a difference, there is a problem in that the actual input value of the data stream reacts sensitively even if there is no significant difference and is incorrectly detected as a concept drift. Therefore, in this paper, in order to reduce such errors, not only the output value of CNN but also the probability vector are used. First, the data entered into the data stream is patterned to learn from the neural network model, and the difference between the output value and probability vector of the current data and the historical data of these learned neural network models is compared to detect the concept drift. The proposed method confirmed that only CNN output values could be used to reduce detection errors compared to how concept drift were detected.

Aiming at properties of remote sensing image data such as high-dimension, nonlinearity and massive unlabeled samples, a kind of probability least squares support vector machine (PLSSVM) classification method based on hybrid entropy and L1 norm was proposed. Firstly, hybrid entropy was designed by combining quasi-entropy with entropy difference, which was used to select the most “valuable” samples to be labeled from massive unlabeled sample set. Secondly, a L1 norm distance measuring was used to further select and remove outliers and redundant data from the sample set to be labeled. Finally, based on originally labeled samples and screened samples, PLSSVM was gained through training. Experimental results on classification of ROSIS hyperspectral remote sensing images show that the overall accuracy and Kappa coefficient of the proposed classification method reach higher accuracy respectively. The proposed method can obtain higher classification accuracy with few training samples, which is much applicable to classification problem of remote sensing images.

We will deal with an n locus model in which mutation and gene conversion are taken into consideration. Also random partitions of the number n determined by chromosomes with n loci should be investigated. The diffusion process describes the time evolution of distributions of the random partitions. In this paper, we find the probability of distribution of the diffusion process with special diffusion operator $L_1$ and we show that the average probability of genes at different loci on one chromosome can be described by the rate of gene frequency of mutation and gene conversion.

Recently, support vector machines (SVMs) are being recognized as competitive tools as compared with other data mining techniques for solving pattern recognition or classification decision problems. Furthermore, many researches, in particular, have proved it more powerful than traditional artificial neural networks (ANNs)(Amendolia et al., 2003; Huang et al., 2004, Huang et al., 2005; Tay and Cao, 2001; Min and Lee, 2005; Shin et al, 2005; Kim, 2003). The classification decision, such as a binary or multi-class decision problem, used by any classifier, i.e. data mining techniques is cost-sensitive. Therefore, it is necessary to convert the output of the classifier into well-calibrated posterior probabilities. However, SVMs basically do not provide such probabilities. So it required to use any method to create probabilities (Platt, 1999; Drish, 2001). This study applies a method to estimate the probability of outputs of SVM to bankruptcy prediction and then suggests credit scoring methods using the estimated probability for bank's loan decision making.

We develop forecast models of daily probabilities of major flares (M- and X-class) based on empirical relationships between photospheric magnetic parameters and daily flaring rates from May 2010 to April 2018. In this study, we consider ten magnetic parameters characterizing size, distribution, and non-potentiality of vector magnetic fields from Solar Dynamics Observatory (SDO)/Helioseismic and Magnetic Imager (HMI) and Geostationary Operational Environmental Satellites (GOES) X-ray flare data. The magnetic parameters are classified into three types: the total unsigned parameters, the total signed parameters, and the mean parameters. We divide the data into two sets chronologically: 70% for training and 30% for testing. The empirical relationships between the parameters and flaring rates are used to predict flare occurrence probabilities for a given magnetic parameter value. Major results of this study are as follows. First, major flare occurrence rates are well correlated with ten parameters having correlation coefficients above 0.85. Second, logarithmic values of flaring rates are well approximated by linear equations. Third, using total unsigned and signed parameters achieved better performance for predicting flares than the mean parameters in terms of verification measures of probabilistic and converted binary forecasts. We conclude that the total quantity of non-potentiality of magnetic fields is crucial for flare forecasting among the magnetic parameters considered in this study. When this model is applied for operational use, it can be used using the data of 21:00 TAI with a slight underestimation of 2-6.3%.

We present the relationship between vector magnetic field parameters and solar major flare occurrence rate. Based on this, we are developing a forecast model of major flare (M and X-class) occurrence rate within a day using hourly vector magnetic field data of Space-weather HMI Active Region Patch (SHARP) from May 2010 to April 2017. In order to reduce the projection effect, we use SHARP data whose longitudes are within ${\pm}60$ degrees. We consider six SHARP magnetic parameters (the total unsigned current helicity, the total photospheric magnetic free energy density, the total unsigned vertical current, the absolute value of the net current helicity, the sum of the net current emanating from each polarity, and the total unsigned magnetic flux) with high F-scores as useful predictors of flaring activity from Bobra and Couvidat (2015). We have considered two cases. In case 1, we have divided the data into two sets separated in chronological order. 75% of the data before a given day are used for setting up a flare model and 25% of the data after that day are used for test. In case 2, the data are divided into two sets every year in order to reduce the solar cycle (SC) phase effect. All magnetic parameters are divided into 100 groups to estimate the corresponding flare occurrence rates. The flare identification is determined by using LMSAL flare locations, giving more numbers of flares than the NGDC flare list. Major results are as follows. First, major flare occurrence rates are well correlated with six magnetic parameters. Second, the occurrence rate ranges from 0.001 to 1 for M and X-class flares. Third, the logarithmic values of flaring rates are well approximated by two linear equations with different slopes: steeper one at lower values and lower one at higher values. Fourth, the sum of the net current emanating from each polarity gives the minimum RMS error between observed flare rates and predicted ones. Fifth, the RMS error for case 2, which is taken to reduce SC phase effect, are smaller than those for case 1.

Probability distribution of the sea surface slope is estimated using sun glitter images derived from visible radiometer on Geostationary Meteorological Satellite (GMS) and surface vector winds observed by spaceborne scatterometers. The brightness of the visible images is converted to the probability of wave surfaces which reflect the sunlight toward GMS in grids of 0.25 deg $\times$ 0.25 deg. Slope and azimuth angle required for the reflection of the sun's ray toward GMS are calculated for each grid from the geometry of GMS observation and location of the sun. The GMS images are then collocated with surface wind data observed by three scatterometers. Using the collocated data set of about 30 million points obtained in a period of 4 years from 1995 to 1999, probability distribution function of the surface slope is estimated as a function of wind speed and azimuth angle relative to the wind direction. Results are compared with those of Cox and Munk (1954a, b). Surface slope estimated by the present method shows narrower distribution and much less directivity relative to the wind direction than that reported by Cox and Munk. It is expected that their data were obtained under conditions of growing wind waves. In general, wind waves are not always developing, and slope distribution might differ from the results of Cox and Munk. Most of our data are obtained in the subtropical seas under clear-sky conditions. This difference of the conditions may be the reason for the difference of slope distribution.

본 논문에서는 전영역 탐색기반의 방법에 비하여 예측화질은 거의 같게 유지하면서 불필요한 계산량을 현저히 줄이는 알고리즘을 제안한다. 제안하는 방법은 움직임 벡터의 확률분포에 따라 탐색패턴을 달리하며, 블록매칭 기준의 비교값을 다르게 함으로써 예측화질을 유지하면서 계산량만 효율적으로 감축할 수 있다. 제안한 알고리즘은 기존의 전영역 탐색 기반인 H.264 PDE 고속 알고리즘과 비교하여 예측화질의 저하가 0~0.02dB이며, 소요된 계산량은 20%~30%정도이다. 제안한 알고리즘은 MPEG-2/4 AVC를 이용하는 실시간 비디오 압축 응용분야에 유용하게 사용될 수 있을 것이다.

In the paper, we propose an algorithm that significantly reduces unnecessary computations, while keeping prediction quality almost similar to that of the full search. In the proposed algorithm, we can reduces only unnecessary computations efficiently by taking different search patterns and error criteria of block matching according to distribution probability of motion vectors. Our algorithm takes only 20~30% in computational amount and has decreased prediction quality about 0~0.02dB compared with the fast full search of the H.264 reference software. Our algorithm will be useful to real-time video coding applications using MPEG-2/4 AVC standards.

최근 몇 년간 SVM(support vector machines)기법은 패턴인식 또는 분류의사결정문제를 위한 분석기법으로서 기존의 데이터마이닝 기법과 비교할 때, 매우 높은 성과를 갖는 것으로 인식되어 왔다. 더 나아나 많은 연구자들은 SVM기법이 1980년대 이후 대표적인 예측 및 분류모형으로 인정받은 인공신경망기법(ANNs : Artificial Neural Networks)에 비해 더 성과가 좋다는 사실을 실증적으로 입증해 왔다(Amendolia et al., 2003; Huang et al., 2004, Huang et al., 2005; Tay and Cao, 2001; Min and Lee, 2005; Shin et al., 2005; Kim, 2003). 일반적으로 이와 같이 다양한 데이터마이닝 기법에 의해 분석되는 이진분류 또는 다분류 의사결정문제들은 특히 금융분야 등에 있어서 오분류비용에 민감하며, 이로 인한 오분류의 경제적 손실도 상대적으로 매우 크다고 할 수 있다. 따라서 기업부도예측모형과 같은 이진분류모형의 결과값을, 부도확률에 기초하여 정교하게 계산된 사후확률의 개념으로서 다분류의 신용등급평가의 문제로 변환할 필요가 있다. 그러나, SVM 모형의 결과값은 기본적으로 그와 같은 부도확률분포를 보여주지 않는다. 따라서, 그러한 확률분포를 정교하게 보여줄 방법을 제시할 필요가 있다(Platt, 1999; Drish, 2001). 본 연구는 AdaBoost 알고리즘기반의 SVM 모형을 이용하여, 이진분류모형으로서 IT 기업의 부실예측모형에 적용한 후, 이 SVM 모형의 예측결과를 SVM의 손실함수에 적용하여 계산된 값을 사후부도확률의 정규분포 특성에 따라 이를 구간화하여 IT기업에 대한 다분류 신용등급 평가의 문제로 전환시키는 방법을 제시하였다. 그리고 본 연구에서 제안하는 방법은 이러한 AdaBoost 알고리즘기반 SVM 모형이 각 기업이 고유한 신용위험(부도확률)을 갖고 있다는 조건하에서, 신용등급부여를 위한 부도확률분포 구간을 정교하게 조정함으로써 오분류 문제를 좀 더 줄일 수 있음을 제시하였다.

Recently, support vector machines (SVMs) are being recognized as competitive tools as compared with other data mining techniques for solving pattern recognition or classification decision problems. Furthermore, many researches, in particular, have proved them more powerful than traditional artificial neural networks (ANNs) (Amendolia et al., 2003; Huang et al., 2004, Huang et al., 2005; Tay and Cao, 2001; Min and Lee, 2005; Shin et al., 2005; Kim, 2003).The classification decision, such as a binary or multi-class decision problem, used by any classifier, i.e. data mining techniques is so cost-sensitive particularly in financial classification problems such as the credit ratings that if the credit ratings are misclassified, a terrible economic loss for investors or financial decision makers may happen. Therefore, it is necessary to convert the outputs of the classifier into wellcalibrated posterior probabilities-based multiclass credit ratings according to the bankruptcy probabilities. However, SVMs basically do not provide such probabilities. So it required to use any method to create the probabilities (Platt, 1999; Drish, 2001). This paper applied AdaBoost algorithm-based support vector machines (SVMs) into a bankruptcy prediction as a binary classification problem for the IT companies in Korea and then performed the multi-class credit ratings of the companies by making a normal distribution shape of posterior bankruptcy probabilities from the loss functions extracted from the SVMs. Our proposed approach also showed that their methods can minimize the misclassification problems by adjusting the credit grade interval ranges on condition that each credit grade for credit loan borrowers has its own credit risk, i.e. bankruptcy probability.