Earticle

본 연구에서는 스마트워치(갤럭시 워치 4–6)에서 얻은 심박수·걸음수 데이터만으로 주 단위 우울(PHQ-9)·불안(GAD-7) 상태를 예측하고, 시간대·행동 패턴 기반 디지털 바이오마커 후보를 탐색하였다. 우울·불안 상태 예측을 위해서 세 가지 그래디언트 부스팅 모델(LightGBM, XGBoost, CatBoost)을 적용하여 비교 평가 하였다. 데이터는 국내 대학생 및 정신건강의학과 외래 환자 18명으로부터 총 68개 주차 표본을 수집하였으며, 1분 단위 시계열에서 시간대, 요일, 활동 세션 특징을 추출해 LightGBM, XGBoost, CatBoost로 분류하였다. 주차 단위 층화 교차검증 기준으로 PHQ-9 3분류 정확도는 약 0.65–0.72 범위로 나타났으며, PHQ-9 3분류에서는 XGBoost의 전체 정확도가 0.721로, GAD-7 이진 분류에서는 LightGBM이 정확도 0.765, 불안 증상군 재현율 0.611로 가장 높은 성능을 보였다. 다만 참여자 단위 교차검증(LOSO)에서는 정확도가 각각 0.647, 0.662로 하락하고 GAD-7에서는 모델 간 우열이 유지되지 않아, 이러한 모델 간 비교 결과는 검증 방식에 민감한 탐색적 결과로, 특정 모델의 일반적 우수성을 의미하지 않는다. 특징 중요도 분석에서는 정오–초저녁(약 10–18시) 시간대의 심박수, 오후 및 저녁 활동량, 심박 변화 비율 및 활동 세션 강도가 주요 예측 특징으로 나타나, 스마트워치 기반 심박·걸음수만으로도 주 단위 정신건강 상태 예측과 해석 가능한 디지털 바이오마커 설계 가능성을 시사한다.

This pilot study aimed to predict weekly depression (PHQ-9) and anxiety (GAD-7) status using only heart rate and step count data obtained from commercial smartwatches (Galaxy Watch 4–6), and to explore candidate digital biomarkers based on time-of-day and behavioral patterns. Specifically, we compared the prediction performance of three gradient boosting models—LightGBM, XGBoost, and CatBoost—for depression and anxiety status classification. We enrolled 18 participants (12 nonclinical undergraduates at Gachon University and 6 psychiatric outpatients at Seoul Asan Medical Center) and constructed 68 participant-week samples, from which we extracted time-of-day, day-of-week, and activity-session features from 1-minute heart rate and step count time series to train gradient boosting classifiers. Under stratified week-level cross-validation, the PHQ-9 three-class models achieved accuracies of approximately 0.65–0.72, with XGBoost showing the highest overall accuracy (0.721), while LightGBM exhibited the strongest overall performance for binary GAD-7 classification with an accuracy of 0.765, a recall of 0.611, and an F1 score of 0.579 for the elevated-anxiety group. However, under leave-one-subject-out(LOSO) validation, accuracies decreased to 0.647 and 0.662, respectively, and the model ranking for GAD-7 was not preserved, indicating that these model comparisons should be interpreted as exploratory baselines sensitive to the validation scheme. Feature importance analyses indicated that time-of-day heart rate (roughly 10:00–18:00), afternoon-to-evening activity, heart-rate change ratios, and activity-session intensity were commonly important predictors, suggesting that smartwatch-based heart rate and step count alone may support weekly mental health status prediction and the design of interpretable digital biomarkers.

본 연구는 종업원 친화적 조직문화가 조직성과에 어떠한 방식으로 반영되는지를 설명가능한 머신러닝 기법을 활용하여 분석한다. 기존 선행연구는 조직문화가 기업의 장기적 성과와 가치 창출에 중요한 역할을 수행함을 이론적으로 강조해 왔으나, 측정의 어려움과 분석기법의 제약으로 인해 문화 요소별 효과를 체계적으로 규명하는 데 한계가 존재하였다. 이에 본 연구는 대규모 기업 리뷰 데이터를 활용하여 종업원 친화적 조직문화를 다차원적으로 측정하고, 트리 기반 앙상블 학습 기법(XGBoost)을 적용하여 조직성과와의 관계를 분석하였다. 특히 SHAP 기법을 활용하여 머신러닝 모형의 예측 결과를 해석함으로써, 각 문화 요소가 조직성과 예측에 기여하는 방향과 크기를 비선형적·조건부 구조 속에서 규명하였다. 분석 결과, 종업원 친화적 조직문화는 회계 기반 성과지표보다는 Tobin’s Q와 같은 시장 기반 성과지표와 보다 밀접하게 연관되는 것으로 나타났으며, 그 효과는 문화 요소 및 평점 구간에 따라 상이한 비선형적 양상을 보였다. 본 연구는 조직문화와 조직성과 간의 관계를 평균적 선형 효과가 아닌 설명력의 구조와 이질성의 관점에서 조명함으로써, 관리회계 분야에서 조직문화 연구의 방법론적·실증적 확장에 기여할 것으로 기대된다.

This study examines how employee-friendly organizational culture is reflected in organizational performance using explainable machine learning techniques. While prior literature has emphasized the importance of organizational culture in shaping long-term firm performance and value creation, empirical evidence identifying which cultural elements matter most remains limited due to measurement challenges and reliance on linear regression approaches. To address these limitations, this study employs large-scale firm-level employee review data to measure employee-friendly organizational culture in a multidimensional manner and applies a tree-based ensemble learning method (XGBoost) to analyze its relationship with organizational performance. Furthermore, the study utilizes Shapley Additive Explanations (SHAP) to enhance model interpretability and to identify the direction and magnitude of each cultural component’s contribution within a nonlinear and conditional framework. The results indicate that employee-friendly organizational culture is more closely associated with market-based performance measures, such as Tobin’s Q, than with accounting-based performance indicators, and that the effects of cultural elements vary across rating levels in a nonlinear manner. By moving beyond average linear effects and focusing on the structure and heterogeneity of explanatory power, this study contributes to the methodological and empirical advancement of organizational culture research in the management accounting literature.

다크넷(Darknet)은 익명성과 보안을 바탕으로 하고 있어 각종 범죄 및 불법 활동에 지속적으로 악용되고 있으며, 이러한 오·남용을 막기 위해 다크넷 트래픽을 정확하게 탐지하고 분류하는 연구는 매우 중요하다. 본 논문에서는 그레디언트 부스팅 기법을 활용한 다크넷 트래픽 탐지 및 분류 기법을 제안하였다. CIC-Darknet2020 데이터셋에 XGBoost와 LightGBM 알고리즘을 적용한 결과, 99.99%의 탐지율과 99% 이상의 분류 성능을 나타내어 기존 연구에 비해 3% 이상 높은 탐지 성능과 13% 이상의 높은 분류 성능을 달성할 수 있었다. 특히, LightGBM 알고리즘의 경우, XGBoost보다 약 1.6배의 학습 시간과 10배의 하이퍼 파라미터 튜닝 실행시간을 단축하여 월등히 우수한 성능으로 다크넷 트래픽 탐지 및 분류를 수행하였다.

Darknet is based on the characteristics of anonymity and security, and this leads darknet to be continuously abused for various crimes and illegal activities. Therefore, it is very important to detect and classify darknet traffic to prevent the misuse and abuse of darknet. This work proposes a novel approach, which uses the Gradient Boosting techniques for darknet traffic detection and classification. XGBoost and LightGBM algorithm achieve detection accuracy of 99.99%, and classification accuracy of over 99%, which could get more than 3% higher detection accuracy and over 13% higher classification accuracy, compared to the previous research. In particular, LightGBM algorithm could detect and classify darknet traffic in a way that is superior to XGBoost by reducing the learning time by about 1.6 times and hyperparameter tuning time by more than 10 times.

High-risk pregnancy endangers both mother and baby, with one maternal death every two minutes in 2023. This study has proposed three Gradient Boosting models?GB-Base, GB-SMOTE, and ZOA-GB?using the West Lombok Pregnancy Risk Prediction Dataset. GB-Base and GB-SMOTE have achieved 90.46% and 90.28% accuracy, while ZOA-GB, using 10 selected features, has reached 88.89%. GB-SMOTE has shown the best performance with an F-score of 84.41%. SHAP has identified Maternal Age, Hemoglobin, and Parity as key features, and DiCE has validated feature-driven prediction control. The study is limited by a single-source dataset, the absence of external-validation, and unexplored optimizers.

NGBoost has shown promising results in probabilistic and point estimation tasks. However, it is vague still whether this method can be scalable to neural architecture system since its base learner is based on decision trees. To resolve this, we design a Neural-NGBoost framework by replacing the base learner with lightweight neural networks and introducing joint gradient estimation for boosting procedure. Based on natural gradient boosting, we iteratively update the neural based learner by inferring natural gradient and update the parameter score with its probabilistic distribution. Experimental results show Neural-NGBoost achieves superior performance across various datasets compared to other boosting methods.

The network security of Plug-and-Charge (PnC) technology in electric vehicle charging systems is typically achieved through the well-known Transport Layer Security (TLS) protocol, which causes high communication overhead. To reduce this overhead, a differential authentication method employing different schemes for individual users has been proposed. However, decisions use a simple threshold approach and no quantitative performance evaluation should be made. In this study, we determined each user's trust using several machine learning algorithms with their charging patterns and compared them. The experimental results reveal that the proposed approach outperforms the conventional approach by 41.36% in terms of round-trip time efficiency, demonstrating its effectiveness in reducing the TLS overhead. In addition, we show the simulation results for three user authentication methods and capture the performance variations under CPU busy waiting scenarios.

A routing strategy based on traffic prediction and dynamic cache allocation for satellite nodes is proposed to address the issues of high propagation delay and overall delay of inter-satellite and satellite-to-ground links in low Earth orbit (LEO) satellite systems. The spatial and temporal correlations of satellite network traffic were analyzed, and the relevant traffic through the target satellite was extracted as raw input for traffic prediction. An improved gradient boosting regression tree algorithm was used for traffic prediction. Based on the traffic prediction results, a dynamic cache allocation routing strategy is proposed. The satellite nodes periodically monitor the traffic load on inter-satellite links (ISLs) and dynamically allocate cache resources for each ISL with neighboring nodes. Simulation results demonstrate that the proposed routing strategy effectively reduces packet loss rate and average end-to-end delay and improves the distribution of services across the entire network.

Prediction of blood glucose (BG) values in type 1 diabetes (T1D) remains an essential and challenging issue. Recently, machine learning methods have been used to solve this problem. We present a forecasting model based on extreme gradient boosting (XGBoost) regression to estimate the BG value of T1D patients in this study. We developed the models using clinical datasets from five T1D patients and forecasted the BG value for the following prediction horizons (PHs) of 10, and 20 minutes. Datasets are divided in two parts, around 60:40 for training and testing. We compared the performance of our proposed model to existing models using several performance metrics, including root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R2). Our suggested XGBoost model outperformed existing models, with average RMSE, MAPE, and R2 of 12.57 mg/dL, 7.15%, 0.94, and 21.93 mg/dL, 13.72%, 0.84 for PH of 10 and 20 minutes, respectively. Finally, the findings of this study are intended to be applied to improve diabetes care.

주가 이동 방향의 정확한 예측이 주식 매매에 관한 전략적 의사결정에 중요한 역할을 할 수 있기 때문에 투자자와 연구자 모두의 관심이 높다. 주가 이동 방향에 관한 기존 연구들을 종합해보면, 주식 시장에 따라서 그리고 예측 기간에 따라서 다양한 변수가 고려되고 있음을 알 수 있다. 이 연구에서는 한국 주식 시장을 대표하는 지수와 주식들을 대상으로 이동 방향 예측 기간에 따라서 어떤 데이터마이닝 기법의 성능이 우수한 것인지를 분석하고자 하였다. 특히, 최근 공개경쟁에서 활발히 사용되며 그 우수성이 입증되고 있는 익스트림 그라디언트 부스팅 기법을 주가 이동 방향 예측 문제에 적용하고자 하였으며, SVM, 랜덤 포리스트, 인공 신경망과 같이 기존 연구에서 우수한 것으로 보고된 데이터마이닝 기법들과 비교하여 분석하였다. 12년간 데이터를 사용하여 1일 후에서 5일 후까지의 이동 방향을 예측하는 실험을 통해서, 예측 기간과 종목에 따라서 선택된 변수들에 차이가 있으며, 1-4일 후 예측에서는 익스트림 그라디언트 부스팅이 다른 기법들과 부분적으로 동등함을 가지면서도 가장 우수함을 확인하였다.

Both investors and researchers are attentive to the prediction of stock price movement directions since the accurate prediction plays an important role in strategic decision making on stock trading. According to previous studies, taken together, one can see that different factors are considered depending on stock markets and prediction periods. This paper aims to analyze what data mining techniques show better performance with some representative index and stock price datasets in the Korea stock market. In particular, extreme gradient boosting technique, proving itself to be the fore-runner through recent open competitions, is applied to the prediction problem. Its performance has been analyzed in comparison with other data mining techniques reported good in the prediction of stock price movement directions such as random forests, support vector machines, and artificial neural networks. Through experiments with the index/price datasets of 12 years, it is identified that the gradient boosting technique is the best in predicting the movement directions after 1 to 4 days with a few partial equivalence to the other techniques.

인터넷 사용의 증가로 인한 외부위협에 대응하는 방안의 하나로서 네트워크 침입탐지를 위한 머신러닝 방식의 연구가 활발하다. 그런데 침입탐지를 위한 머신러닝 알고리즘의 적용에 있어서 소수 클래스로 인해 발생하는 데이터 불균형 문제가 지적된다. 네트워크 침입탐지를 비롯한 일반적인 분류에서는 데이터 불균형에 의한 문제점보다는 대개 전체 모형의 정확도가 목표인 경우가 많아서, 데이터 불균형에 대한 대처가 쉽지 않을 수 있다. 네트워크 침입탐지에 대한 랜덤포레스트(RF) 모형에서 데이터 불균형 문제가 있음을 본 논문에서 확인하 고, 이를 개선하기 위한 그래디언트부스팅 분류기(GBC)의 구성 및 효과를 정리하였다. 랜덤포레스트 모형과 그래디언트부스팅 분류기의 구성은 KDDTrain+ 데이터를 활용하였고, KDDTest+ 데이터로써 구성한 모형을 평가하였다. RF모형과 GBC의 성능에서 보이는 차이점 은 빈도가 낮은 침입유형에서 RF모형에 비해 정확도의 큰 변화 없이 GBC의 정밀도 및 재현율이 높은 것을 확인하였다. GBC의 이러한 효과는 특히 피해가 큰 침입유형에 대해 높은 탐지율로써 전체 피해를 줄이는 효과를 기대할 수 있다.

As a way to respond to external threats due to the increase of Internet usage, researches on machine learning methods for network intrusion detection becomes more active. However the problem of data imbalance caused by minority classes is pointed out in the application of machine learning algorithms for intrusion detection. In general classification problems including network intrusion detection, the accuracy of the entire model is often the goal rather than handling the problems caused by such minority classes, so it may not be easy to deal with data imbalance. In this paper, we checked that there is a data imbalance problem in the random forest model used in network intrusion detection, and organized the composition and effect of the gradient boosting classifier for this point. The Random Forest (RF) model and the Gradient Boosting Classifier (GBC) were constructed using the KDDTrain+ data and evaluated using the KDDTest+ data. The difference in the performance of the RF model and the GBC is that the precision and recall of the GBC are higher than that of the RF model without a significant change in accuracy in low-frequency intrusion types. This effect of GBC is expected to have the effect of reducing the overall damage by detecting intrusion types that cause particularly serious damage with a higher probability.

본 연구에서 의사결정트리는 지니 불순도(Gini Impurity)를 기준으로 데이터를 재귀적으로 분할하여 규칙 기반 예측 수행한다. 랜덤포레스트는 다수의 의사결정트리를 무작위 샘플링을 통해 학습시키고 투표 방식을 통해 최종 예측 수행한다. 그래디언트부스팅은 약한 분류기를 순차적으로 학습시키며, 이전 단계의 오차를 줄여나가는 방식으로 예측 성능 개선한다. 또한 해석 용이성을 위해 학습은 depth=3으로 고정했다. 학습이 더 깊어도 세 가지 모델들의 구조적인 그림은 max_depth=3까지만 나타냈다. 랜덤포레스트의 최종 예측 트리는 숲 전체의 다수결 판단을 근사하는 설명용 트리이다. 다시 말해서 이들의 최종 예측 트리는 숲의 수백 트리들을 한 장의 트리로 보여줄 수 없으므로 일반적으로 대리 트리 기법을 사용하였다. 그래디언트부스팅 트리에서는 300번째 트리를 출력했다. 인공지능 모델의 성능을 평가하는 데 중요한 지표들인 정확도, 정밀도, 재현율, 그리고 F1 점수 개념이다. 본 연구에서도 각각의 모델에 대해 정확도, 정밀도, 재현율, 그리고 F1 점수를 기반으로 성능을 평가하였다. 또한 각각의 모델에서 도출된 변수 중요도를 비교하여 해석적 차이를 분석하였다.

In this paper Rule-based prediction is performed by recursively dividing data based on Gini Impurity in the decision tree. Random forest trains multiple decision trees through random sampling and performs final prediction through voting. Gradient boosting learns weak classifiers sequentially and improves prediction performance by reducing errors in previous steps. Also for ease of interpretation, the learning is fixed as depth=3. Even if the learning is deeper, the structural picture of the three models is shown only up to max_depth=3. The final prediction tree of the random forest is an explanatory tree that approximates the majority judgment of the entire forest. In other words, their final prediction tree cannot show hundreds of trees in a forest as a single tree, so they generally used a surrogate tree technique. In the gradient boosting tree, the 300th tree was output. The concepts of Accuracy, Precision, Recall, and F1-Score are important indicators for evaluating the performance of artificial intelligence models. In this study, the performance of each model was evaluated based on Accuracy, Precision, Recall, and F1 score. In addition, analytical differences were analyzed by comparing the variable importance derived from each model.

In this study, we propose a technique to automatically generate transfer documents using sensor data from livestock manure transfer systems. The research involves analyzing sensor data and applying machine learning techniques to derive optimized outcomes for livestock manure transfer documents. By comparing and contrasting with existing documents, we present a method for automatic document generation. Specifically, we propose the utilization of Gradient Boosting, a machine learning algorithm. The objective of this research is to enhance the efficiency of livestock manure and liquid byproduct management. Currently, stakeholders including producers, transporters, and processors manually input data into the livestock manure transfer management system during the disposal of manure and liquid byproducts. This manual process consumes additional labor, leads to data inconsistency, and complicates the management of distribution and treatment. Therefore, the aim of this study is to leverage data to automatically generate transfer documents, thereby increasing the efficiency of livestock manure and liquid byproduct management. By utilizing sensor data from livestock manure and liquid byproduct transport vehicles and employing machine learning algorithms, we establish a system that automates the validation of transfer documents, reducing the burden on producers, transporters, and processors. This efficient management system is anticipated to create a transparent environment for the distribution and treatment of livestock manure and liquid byproducts.

In this study, the machine learning which has been widely used in prediction algorithms recently was used. the research point was the CD(chudong) point which was a representative point of Daecheong Lake. Chlorophyll-a(Chl-a) concentration was used as a target variable for algae prediction. to predict the Chl-a concentration, a data set of water quality and quantity factors was consisted. we performed algorithms about random forest and gradient boosting with Python. to perform the algorithms, at first the correlation analysis between Chl-a and water quality and quantity data was studied. we extracted ten factors of high importance for water quality and quantity data. as a result of the algorithm performance index, the gradient boosting showed that RMSE was 2.72 mg/m³ and MSE was 7.40 mg/m³ and R² was 0.66. as a result of the residual analysis, the analysis result of gradient boosting was excellent. as a result of the algorithm execution, the gradient boosting algorithm was excellent. the gradient boosting algorithm was also excellent with 2.44 mg/m³ of RMSE in the machine learning hyperparameter adjustment result.

본 논문에서는 그래디언트 부스팅 모형을 활용하여 정부의 중소기업 연구개발 지원 결정에 영향을 미치는 요인들을 파악하였다. 기존 연구가 사후적으로 정부의 연구개발 지원이 수혜 기업에 미친 영향을 분석하는 것에 중점을 두었다면, 본 논문은 정부의 연구개발 지원 결정 방식을 파악하고, 그 방식이 기업에게 제공하는 유인을 분석하고자 하였다. 이를 위하여 본 논문은 지원금 결정에 영향을 미치는 다양한 잠재적 요인들을 선택하고, 기계학습 접근법을 활용하여 추정오차 축소효과가 큰 요인들을 선별하였다. 구체적으로 본 논문은 한국과학기술평가원이 구축한 국가연구개발조사분석 자료와 한국신용평가자료를 연결한 자료에 그래디언트 부스팅(Gradient Boosting) 모형을 적용하여 지원금 추정모형을 구축하였다. 본 논문에서 구축한 그래디언트 부스팅 모형은 선형회귀분석 응용모형에 비해 평균제곱근오차를 7.20% 축소할 수 있었다. 각 변수의 순열 중요도(permutation importance)를 분석한 결과 연구성과지표 및 연구개발비가 추정오차 축소에 기여가 큰 것으로 파악되었다. 그리고 각 변수의 부분의존도(Partial Dependence Plot: PDP) 및 SHAP 값(SHAP value: SHapley Additive exPlanation value)을 분석한 결과 연구성과지표가 좋고 연구개발비 지출이 큰 기업이 많은 연구개발 지원금을 받는 반면, 영업이익이 크고 자기자본회전율이 높은 기업은 적은 지원금을 받는 경향이 발견되었다. 본 연구의 결과는 현재 중소기업 연구개발 지원금 배분 방식이 연구성과지표 제고 및 연구개발투자 증가 유인은 제공하나, 기업 경영성과 제고 유인은 취약함을 시사한다.

In this paper, we build a gradient Boosting model to predict government SME R&D subsidy, select features of high importance, and measure the impact of each features to the predicted subsidy using PDP and SHAP value. Unlike previous empirical researches, we focus on the effect of the R&D subsidy distribution pattern to the incentive of the firms participating subsidy competition. We used the firm data constructed by KISTEP linking government R&D subsidy record with financial statements provided by NICE, and applied a Gradient Boosting model to predict R&D subsidy. We found that firms with higher R&D performance and larger R&D investment tend to have higher R&D subsidies, but firms with higher operation profit or total asset turnover rate tend to have lower R&D subsidies. Our results suggest that current government R&D subsidy distribution pattern provides incentive to improve R&D project performance, but not business performance.

This paper proposes an efficient data-driven approach to build models for predicting energy consumption in buildings. Data used in this research is collected by installing humidity and temperature sensors at different locations in a building. In addition to this, weather data from nearby weather station is also included in the dataset to study the impact of weather conditions on energy consumption. One of the main emphasize of this research is to make feature selection independent of domain knowledge. Therefore, to extract useful features from data, two different approaches are tested: one is feature selection through principal component analysis and second is relative importance-based feature selection in original domain. The regression model used in this research is gradient boosting regression and its optimal parameters are chosen through a two staged coarse-fine search approach. In order to evaluate the performance of model, different performance evaluation metrics like r2-score and root mean squared error are used. Results have shown that best performance is achieved, when relative importance-based feature selection is used with gradient boosting regressor. Results of proposed technique has also outperformed the results of support vector machines and neural network-based approaches tested on the same dataset.

This study examines the satisfaction effect of classics learn course using deep learning Ensemble models such as Voting, Random Forest, and Gradient Boosting models. The following were the main findings of the comparison of their predictive abilities. First, in the classification predictions of the necessity of classics learning courses, the Ensemble models were sensitive to hyperparameters but generally provided higher R-squared values compared to individual machine learning models. Second, in the regression predictions of satisfaction, the Voting ensemble model showed a higher R-squared and lower RMSE compared to individual deep learning models. Third, in the Random Forest model, which made it the best model with the least risk of overfitting if we chose the appropriate hyperparameters. Fourth, for the Gradient Boosting model, by increasing the learning rate to 0.2 the R-squared values for the evaluation dataset were higher, and the RMSE was lower, indicating that this setup led to better predictions, showing that the necessity of classics learning course increased after the course.

학생들의 중도 탈락은 대학의 재정적 손실 뿐 아니라, 학생 개개인 및 사회적으로도 부정적인 영향을 끼친다. 이러한 문제를 해결하기 위해 기계 학습을 이용하여 대학생들의 중도 탈락 여부를 예측하고자 하는 다양한 시도가 이루어지고 있다. 본 논문에서는 대학생들의 중도 탈락 여부를 예측하기 위해 DNN(Deep Neural Network)과 LGBM(Light Gradient Boosting Machine)을 이용한 모델을 구현하고 성능을 비교하였다. 학습 데이터로는 서울 소재 중소규모 4년제 대학인 A 대학의 20,050명의 학생을 대상으로 수집된 학적 및 성적 데이터를 학습에 이용하였다. 원본 데이터의 140여개의 속성 중 중도 탈락 여부를 나타내는 속성과의 상관계수가 0.1 이상인 속성들만 추출하여 학습하였다. 두 모델의 성능 실험 결과, DNN과 LGBM의 F1-스코어는 0.798과 0.826이었으며, LGBM이 DNN에 비해 2.5% 나은 예측 성능을 보였다.

Dropouts of students not only cause financial loss to the university, but also have negative impacts on individual students and society together. To resolve this issue, various studies have been conducted to predict student dropout using machine learning. This paper presents a model implemented using DNN (Deep Neural Network) and LGBM (Light Gradient Boosting Machine) to predict dropout of university students and compares their performance. The academic record and grade data collected from 20,050 students at A University, a small and medium-sized 4-year university in Seoul, were used for learning. Among the 140 attributes of the collected data, only the attributes with a correlation coefficient of 0.1 or higher with the attribute indicating dropout were extracted and used for learning. As learning algorithms, DNN (Deep Neural Network) and LightGBM (Light Gradient Boosting Machine) were used. Our experimental results showed that the F1-scores of DNN and LGBM were 0.798 and 0.826, respectively, indicating that LGBM provided 2.5% better prediction performance than DNN.

본 연구에서는 그래디언트 부스팅 기계학습과 유전 알고리즘을 이용하여 일별 주식 포트폴리오를 동적으로 구성하는 시스템을 구축하고 트레이딩 시뮬레이션을 통해 성능을 분석하였다. 이를 위해 유가증권시장과 코스닥시장에 상장된 종목들의 가격 데이터 및 투자자별 거래정보를 포함한 다양한 데이터를 수집하고, 전처리 과정과 변수가공을 통해 학습-예측에 이용될 변수들을 생성하였다. 첫 번째 실험에서는 예측정확도와 정밀도, 재현율 및 F1 점수 등 네 가지 지표를 활용하여 그래디언트 부스팅 기법들(XGBoost, LightGBM, CatBoost)의 성능을 비교 평가하였다. 두 번째 실험에서는 전 단계에서 선택된 LightGBM과 유전 알고리즘을 적용하여 상장 종목들의 일별 수익 여부를 학습-예측하였다. 그리고 예측된 수익 발생확률을 바탕으로 종목을 선별하여 트레이딩 시뮬레이션을 시행하고, CAGR, MDD, 사프지수 및 변동성 측면에서 코스피, 코스닥 지수와의 성능을 비교 평가하였다. 분석 결과, 제안된 전략들 모두 네 가지 성능평가 지표상에서 시장 평균을 넘어서는 것으로 나타났으며, 그래디언트 부스팅과 유전 알고리즘의 결합이 주식 가격 예측에 효과적으로 이용될 수 있음을 보여주었다.

In this study, we developed a system to dynamically balance a daily stock portfolio and performed trading simulations using gradient boosting and genetic algorithms. We collected various stock market data from stocks listed on the KOSPI and KOSDAQ markets, including investor-specific transaction data. Subsequently, we indexed the data as a preprocessing step, and used feature engineering to modify and generate variables for training. First, we experimentally compared the performance of three popular gradient boosting algorithms in terms of accuracy, precision, recall, and F1-score, including XGBoost, LightGBM, and CatBoost. Based on the results, in a second experiment, we used a LightGBM model trained on the collected data along with genetic algorithms to predict and select stocks with a high daily probability of profit. We also conducted simulations of trading during the period of the testing data to analyze the performance of the proposed approach compared with the KOSPI and KOSDAQ indices in terms of the CAGR (Compound Annual Growth Rate), MDD (Maximum Draw Down), Sharpe ratio, and volatility. The results showed that the proposed strategies outperformed those employed by the Korean stock market in terms of all performance metrics. Moreover, our proposed LightGBM model with a genetic algorithm exhibited competitive performance in predicting stock price movements.