PURPOSES : In this study, model-agnostic methods are applied for interpreting machine learning models, such as the feature global effect, the importance of a feature, the joint effects of features, and explaining individual predictions. METHODS : Model-agnostic global interpretation techniques, such as partial dependence plot (PDP), accumulated local effect (ALE), feature interaction (H-statistics), and permutation feature importance, were applied to describe the average behavior of a machine learning model. Moreover, local model-agnostic interpretation methods, individual conditional expectation curves (ICE), local surrogate models (LIME), and Shapley values were used to explain individual predictions. RESULTS : As global interpretations, PDP and ALE-Plot demonstrated the relationship between a feature and the prediction of a machine learning model, where the feature interaction estimated whether one feature depended on the other feature, and the permutation feature importance measured the importance of a feature. For local interpretations, ICE exhibited how changing a feature changes the interested instance’s prediction, LIME explained the relationship between a feature and the instance’s prediction by replacing the machine model with a locally interpretable model, and Shapley values presented how to fairly contribute to the instance’s prediction among the features. CONCLUSIONS : Model-agnostic methods contribute to understanding the general relationship between features and a prediction or debut a model from the global and/or local perspective, securing the reliability of the learning model.

In this study, the positions of Cs-137 gamma ray source are estimated from the plastic scintillating fiber bundle sensor with length of 5 m, using machine learning data analysis. Seven strands of plastic scintillating fibers are bundled by black shrink tube and two photomultiplier tubes are used as a gamma ray sensing and light measuring devices, respectively. The dose rate of Cs-137 used in this study is 6 μSv·h−1. For the machine learning modeling, Keras framework in a Python environment is used. The algorithm chosen to construct machine learning model is regression with 15,000 number of nodes in each hidden layer. The pulse-shaped signals measured by photomultiplier tubes are saved as discrete digits and each pulse data consists of 1,024 number of them. Measurements are conducted separately to create machine learning data used in training and test processes. Measurement times were different for obtaining training and test data which were 1 minute and 5 seconds, respectively. It is because sufficient number of data are needed in case of training data, while the measurement time of test data implies the actual measuring time. The machine learning model is designated to estimate the source positions using the information about time difference of the pulses which are created simultaneously by the interaction of gamma ray and plastic scintillating fiber sensor. To evaluate whether the double-trained machine learning model shows enhancement in accuracy of source position estimation, the reference model is constructed using training data with one-time learning process. The double-trained machine learning model is designed to construct first model and create a second training data using the training error and predetermined coefficient. The second training data are used to construct a final model. Both reference model and double-trained models constructed with different coefficients are evaluated with test data. The evaluation result shows that the average values calculated for all measured position in each model are different from 7.21 to 1.44 cm. As a result, by constructing the double-trained machine learning model, the final accuracy shows 80% of improvement ratio. Further study will be conducted to evaluate whether the double-trained machine learning model is applicable to other data obtained from measurement of gamma ray sources with different energy and set a methodology to find optimal coefficient.

Algal bloom is an ongoing issue in the management of freshwater systems for drinking water supply, and the chlorophyll-a concentration is commonly used to represent the status of algal bloom. Thus, the prediction of chlorophyll-a concentration is essential for the proper management of water quality. However, the chlorophyll-a concentration is affected by various water quality and environmental factors, so the prediction of its concentration is not an easy task. In recent years, many advanced machine learning algorithms have increasingly been used for the development of surrogate models to prediction the chlorophyll-a concentration in freshwater systems such as rivers or reservoirs. This study used a light gradient boosting machine(LightGBM), a gradient boosting decision tree algorithm, to develop an ensemble machine learning model to predict chlorophyll-a concentration. The field water quality data observed at Daecheong Lake, obtained from the real-time water information system in Korea, were used for the development of the model. The data include temperature, pH, electric conductivity, dissolved oxygen, total organic carbon, total nitrogen, total phosphorus, and chlorophyll-a. First, a LightGBM model was developed to predict the chlorophyll-a concentration by using the other seven items as independent input variables. Second, the time-lagged values of all the input variables were added as input variables to understand the effect of time lag of input variables on model performance. The time lag (i) ranges from 1 to 50 days. The model performance was evaluated using three indices, root mean squared error-observation standard deviation ration (RSR), Nash-Sutcliffe coefficient of efficiency (NSE) and mean absolute error (MAE). The model showed the best performance by adding a dataset with a one-day time lag (i=1) where RSR, NSE, and MAE were 0.359, 0.871 and 1.510, respectively. The improvement of model performance was observed when a dataset with a time lag up of about 15 days (i=15) was added.

본 연구는 기계학습을 통한 수량예측모델을 이용하여 이상기상에 따른 WCM의 DMY 피해량을 산출하기 위한 목적으로 수행하였다. 수량예측모델은 WCM 데이터 및 기상 데이터를 수집 후 가공하여 8가지 기계학습을 통해 제작하였으며 실험지역은 경기도로 선정하였다. 수량예측모델은 기계학습 기법 중 정확성이 가장 높은 DeepCrossing (R2=0.5442, RMSE=0.1769) 기법을 통해 제작하였다. 피해량은 정상기상 및 이상기상의 DMY 예측값 간 차이로 산출하였다. 정상기상에서 WCM의 DMY 예측값은 지역에 따라 차이가 있으나 15,003~17,517 kg/ha 범위로 나타났다. 이상기온, 이상강수량 및 이상풍속에서 WCM의 DMY 예측 값은 지역 및 각 이상기상 수준에 따라 차이가 있었으며 각각 14,947~17,571 kg/ha, 14,986~17,525 kg/ha 및 14,920~17,557 kg/ha 범위로 나타났다. 이상기온, 이상강수량 및 이상풍속에서 WCM의 피해량은 각각 –68~89 kg/ha, -17~17 kg/ha 및 – 112~121 kg/ha 범위로 피해로 판단할 수 없는 수준이었다. WCM의 정확한 피해량을 산출하기 위해서는 수량예측모델에 이용하는 이상기상 데이터 수의 증가가 필요하다.

목적 : 인공지능의 기계학습 또는 심층학습을 이용한 연구가 다양한 분야에서 시도되고 있다. 본 연구는 공공 시력데이터를 자동화 수집하고, 수집한 데이터를 기계학습에 적용 및 예측하였다. 다양한 학습모델간 성능을 비교 함으로써, 시과학분야에서 적용 가능한 기계학습 최적화모델을 제시함에 있다. 방법 : 국민건강보험(NHISS) 및 통계포털(KOSIS)에 발표된 국민 시력분포 현황관련 자료를 특정 색인을 포함하 는 자료검색기법인 크롤링(crawling)을 사용하여 검색 및 수집을 자동화하였다. 2011년부터 2018년까지 보고된 모든 자료를 수집하였으며, 데이터 학습을 위해 Linear Regression, LASSO, Ridge, Elastic Net, Huber Regression, LASSO/LARS, Passive Aggressive Regressor 그리고 Pansacregressor 총 8개 모델을 사용하여 각각 데이터 학습 하였다. 결과 : 수집한 데이터를 기반으로 기계학습 모델을 통해 2018년을 예측하였다. 각 모델간 2018년도 실제-예측데 이터 차이를 MAE(Mean Absolute Error)와 RMSE(Root Mean Square Error) 점수로 각각 나타냈다. 학습모델 별 차이 중 MAE 평가결과 모델간 우/좌 Linear Regression(0.22/0.22), LASSO(0.83/0.81), RIDGE(0.31/0.31), Elastic Net(0.86/0.84), Huber Regression(0.14/0.07), LASSO/LARS(0.15/0.14), Passive Aggressive Regressor (0.29/0.18) 그리고 RANSA Regressor(0.22/0.22)를 보였다. RMSE에서 Linear Regression(0.40/0.40), LASSO (1.08/1.06), Ridge(0.54/0.54), Elastic Net(1.19/1.17), Huber Regression(0.20/0.20), LASSO/LARS(0.24/0.23), Passive Aggressive Regressor(0.21/0.58) 그리고 RANSA Regressor(0.40/0.40) 각각 나타냈다. 결론 : 본 연구는 자동화 자료검색 및 수집을 위한 크롤링 기법을 이용하여 데이터를 수집하였다. 이를 기반으 로 고전 선형모델을 기계학습에 적용할 수 있도록 하고, 데이터 학습을 위한 8개 학습모델들 간 성능을 비교하였다.

It is known that the growth and development of the mosquito are greatly affected by the change of the meteorological factors. In particular, temperature and precipitation are closely related to the life cycle of the mosquito, and their effects have different characteristics for each species of mosquito. Therefore, to develop a mosquito activity index based on mosquito density, it is essential to develop a prediction model based on weather data. In this study, we developed a functional formula that can estimate the change of mosquito density according to the change of meteorological factors using the mosquito classification data of Incheon region collected from 2011 to 2017. Also, using the data of the digital mosquito monitoring system (DMS) from April to October 2018, mosquito activity index according to characteristics of space in city was developed. In order to reflect the temporal characteristics of the mosquito life-cycle, we used data of temperature and precipitation prior to 1-2 weeks, and used land cover data to reflect the spatial characteristics of mosquito density. Density of Culex pipiens collected in the Incheon area were gradually increased when the average temperature increased two weeks ago after adjusting the precipitation. However, when the average temperature reached 22°C, the density of Culex pipiens marked a peak, and above the 22°C, the density was decreased. The predicted mosquito activity index calculated by applying the machine learning method to the DMS data of the Incheon area is designed to calculate from 1 to 10 grades. The accuracy of the mosquito activity index was 87% when the one grade error was allowed.

한국형 e-Navigation의 내항성 안전 모듈은 운항 중인 선박을 실시간으로 모니터링하고 내항성의 이상 상태를 사전에 경고함으로써 선박의 안정성을 확보하는 선내 원격 모니터링 서비스 중 하나이다. 일반적으로 선박설계를 위한 내항성능은 주어진 조건에서 선체 운동 시뮬레이션을 수행하여 평가하여 왔다. 하지만 운항 중 선박의 내항성능을 실시간으로 평가하기 위해 이러한 시뮬레이션을 실제 운항조건에 맞추어 수행하는 것은 계산시간의 한계로 인해 현실적이지 않다. 본 연구에서는 기계학습 기반의 근사모델을 활용하여 선박의 내항성능 평가 요소들 중 하나인 횡동요 운동특성을 합리적으로 보다 빠르게 예측하는 방법을 소개하고자 한다. 다양한 학습 기법과 데이터의 샘플링 조건을 적용하여, 얻어진 근사모델의 결과와 운동해석 결과의 오차가 거의 1% 내로 일치함을 보였다. 따라서 이러한 방법을 활용하면 선박의 실시간 내항 성능을 평가하는데 효율적으로 사용할 수 있을 것으로 판단된다.

A novel disaggregation model that combines a machine learning model and kriging of residuals is presented to map precipitation at a fine scale from coarse scale precipitation data. Random forest (RF) and fine scale auxiliary variables are used to estimate trend components at a fine scale. Residual components are then estimated by area-to-point residual kriging. A case study of spatial disaggregation of TRMM monthly precipitation data acquired over the Korean peninsula is carried out to illustrate the potential of the presented disaggregation method. From the evaluation results, the presented method outperformed the RF-based disaggregation method that only considers trend components and ignores residual components, in terms of accuracy statistics and the ability of coherent predictions. This case study indicates that accounting for residual components by applying a proper spatial prediction method such as area-to-point kriging is very important in spatial disaggregation of coarse scale spatial data, even though advanced regression models such as RF could have high goodness of fit for the quantification of relationships between a target attribute and auxiliary variables.