Earticle

This study examined associations between diverse accessibility indicators and population and business density using Extreme Gradient Boosting Decision Tree Regressor (XGB) and Interpretable Machine Learning (XAI). The main results are as follows. First, the results of feature importance reveal that the accessibility indicators exerted a more considerable contribution to predicting the density than control variables. Specifically, accessibility to hospitals appeared to have the most significant impact on explaining population and business density. In addition, accessibility to elementary, middle, and high schools showed high importance in explaining the population density, while it was found to have a smaller importance in explaining the business density. Second, findings of partial dependence plots derived non-linear relationships between accessibility and density. For instance, the population or business density was highest when the travel time to elementary, middle, and high school was between 5 and 7 minutes. This study contributes to (1) offering a more detailed understanding of the relationship between transportation accessibility and density and (2) providing associated discussions and policy implications.

This study proposes an interpretable framework for corporate bankruptcy prediction by integrating machine learning and explainable artificial intelligence (XAI) techniques. While prior bankruptcy prediction studies have primarily focused on improving predictive accuracy using static financial indicators, limited attention has been given to incorporating dynamic financial trends and translating model interpretation results into practical decision- making frameworks. To address this limitation, this study applies a flatten strategy that transforms recent three-year financial data into a structured vector representation while preserving temporal information. Based on this approach, an XGBoost model was developed using financial statement variables, financial ratios, growth indicators, and delta variables reflecting year-over-year changes. To address the severe class imbalance problem inherent in bankruptcy prediction, repeated undersampling experiments were conducted, and model performance was evaluated using Recall and G-Mean, which are particularly suited for imbalanced classification tasks. The empirical results show that the dataset combining static and dynamic financial variables achieved the best predictive performance, particularly in Recall and G-Mean. SHAP analysis further revealed that market value, liquidity, profitability, and capital structure variables play critical roles in bankruptcy prediction. In particular, delta variables related to profitability and growth trends demonstrated high explanatory power, suggesting that corporate bankruptcy is more closely associated with the deterioration trajectory of financial conditions rather than a snapshot of financial status at a single point in time. Based on the SHAP analysis results, this study systematizes bankruptcy-related financial variables into five interpretive categories: firm size, liquidity, profitability, capital structure, and financial trend. This categorization provides a practical interpretation framework that goes beyond binary bankruptcy prediction, explaining through which financial dimensions risk emerges and supporting financial risk diagnosis and decision-making. This study contributes to the literature by integrating temporal financial trends with interpretable machine learning and by proposing a practical SHAP-based financial risk interpretation framework for corporate bankruptcy analysis.

본 연구는 블랙박스 머신러닝 모형의 예측 메커니즘을 해석하기 위한 도구로서 부분 의존성(PD) 플롯과 누적 국소 효과(ALE) 플롯의 원리와 적용 방법을 소개하고 실증 분석 사례를 제시하였다. 1차원 플롯과 2차원 플롯이 각각 주효과와 상호작용효과를 어떻게 드러내는지 확인하고, 변수 간 높은 상관관계가 있을 때 PD와 ALE의 장단점을 비교하였다. 또한 ALE 기반의 정량적 효과 크기 지표와 통계적 유의성 평가 절차를 소개하였다. 국민건강영양조사 자료를 활용하여 성인의 건강 관련 삶의 질(HINT-8)을 예측하는 랜덤 포레스트 모형을 구축하고, 이 앙상블 모형에 PD 플롯과 ALE 플롯을 적용하여 복잡한 예측 메커니즘을 해석하였다. 분석 결과, 가장 중요한 예측변수인 범불안장애 점수는 건강 관련 삶의 질에 뚜렷한 부정적 영향을 보였으며, 1차원 플롯을 통해 효과 크기가 점수 증가에 따라 점차 감소하는 비선형적 양상을 확인할 수 있었다. 2차원 플롯 분석에서는 가장 큰 상호작용효과를 보인 변수 쌍의 하나인 연령과 주관적 건강 인식 간의 복잡한 이원 상호작용 패턴을 시각화할 수 있었다. 본 연구는 예측 성능이 우수한 블랙박스 모형의 해석 가능성을 높이는 실용적 방법을 제시하였다는 점에서 의의가 있다.

This study introduces Partial Dependence (PD) and Accumulated Local Effects (ALE) plots as interpretable tools for black-box machine learning models and demonstrates their application through empirical analysis. We illustrate how one-dimensional and two-dimensional plots capture main effects and interaction effects respectively, and compare the relative merits of PD and ALE plots when dealing with correlated predictors. We also describe ALE-based effect size indices and significance testing procedures. We applied these methods to a random forest model predicting adult health-related quality of life (HINT-8) using Korea National Health and Nutrition Examination Survey data. The generalized anxiety disorder score emerged as the most influential predictor, exhibiting a strong negative effect on quality of life. One-dimensional plots revealed a nonlinear pattern with diminishing marginal effects at higher scores. Two-dimensional plots uncovered a complex interaction between age and subjective health perception, representing one of the strongest interaction effects in the model. Our findings illustrate how PD and ALE plots can make sophisticated black-box models accessible and actionable for applied researchers.

최근 늘어나고 있는 이상 기상 현상으로 산사태 위험이 점차 증가하고 있다. 산사태는 막대한 인명 피해와 재산 피해를초래할 수 있기에 이러한 위험을 사전에 평가함은 매우 중요하다. 최근 기술 발전으로 인해 능동형 원격탐사 방법을 사용하여더 정확하고 상세한 지표 변위 및 강수 데이터를 얻을 수 있게 되었다. 그러나 이러한 데이터를 활용하여 산사태 예측 모델을개발하는 연구는 찾기 힘들다. 따라서 본 연구에서는 합성개구레이더 간섭법(InSAR)을 사용한 지표 변위 자료와 하이브리드고도면 강우(HSR) 추정 기법을 통한 강수 정보를 활용하여 산사태 민감도를 예측하는 기계학습 모델을 제시하고 있다. 나아가기계학습의 블랙박스 문제를 극복할 수 있는 해석가능한 기계학습 방법인 SHAP을 이용하여 산사태 민감도의 영향 변수에대한 중요도를 체계적으로 평가하였다. 경상북도 울진군을 대상으로 사례 연구를 수행한 결과, XGBoost가 가장 좋은 예측성능을 보이며, 도로로부터의 거리, 지표 고도, 일 최대 강우 강도, 48시간 선행 누적 강우량, 사면 경사, 지형습윤지수, 단층으로부터의 거리, 경사도, 지표 변위, 하천으로부터의 거리가 산사태 예측에 영향을 미치는 주요 변수로 밝혀졌다. 특히, 능동형원격탐사를 통해 얻은 자료인 강우 강도와 지표 변위의 절댓값이 높을수록 산사태 발생 확률이 높음을 확인하였다. 본 연구는능동형 원격탐사 자료의 산사태 민감도 연구에서의 활용 가능성을 실증적으로 보여주고 있으며, 해당 자료를 바탕으로 시공간적으로 변하는 산사태 민감도를 도출함으로써 향후 산사태 민감도 모니터링에 효과적으로 활용될 수 있을 것으로 기대된다.

In recent years, increasing extreme weather events have further escalated the risk of landslides. Since the landslide leads to serious loss of human life and property, it is very important to evaluate such risks in advance. The recent advancements in active remote sensing methods have made it easier to access more accurate and detailed surface displacement and rainfall data. However, there have been few works of using such actively remote sensed datasets. Therefore, this study attempts to suggest a machine learning model to predict landslide susceptibility using InSAR (Interferometric SAR) displacement data and HSR (Hybrid Surface Rainfall) estimates as inputs for machine learning models. Moreover, we evaluate the influence of predictor variables for landslide susceptibility systematically by utilizing a SHAP (SHapley Additive exPlanations) model as an interpretable machine learning approach, which has the strength of overcoming a black-box problem with general machine learning models. As a result of case study for Uljin-gun, Gyeongsangbuk-do, XGBoost shows the best predictive performance for landslides and predictor variables such as distance from roads, elevation, maximum daily rainfall intensity, vegetation index, 48 hours accumulated antecedent rainfall, slope, terrain wetness, distance from fault-line, absolute value of surface displacement, and distance from rivers on landslide prediction are identified as key factors for influencing landslide prediction. In particular, rainfall intensity and the absolute value of surface displacement obtained from active remote sensing are contributed to a higher probability of landslide occurrence. This study would be significant in the sense that it presents the applicability of active remote sensed data for landslide susceptibility analysis through the empirical examination. Also, spatially and temporally changing landslide susceptibility based on such actively sensed data could be effectively utilized for landslide monitoring.

본 연구는 기계학습을 기반으로 회계부정탐지모형을 개발하는 데 있어 내부회계관리제도 운영보고서의 텍스트 감성지수가 추가적 예측기여도를 갖는지 탐색적으로 검증한 후 개발된 모형의 해석을 시도하였다. 회계부정의 대용변수로는 금융감독원의 감리지적사례를 사용하였고, 기계학습에 사용될 데이터는 재무비율과 내부회계관리제도 변수 등 총 183개 의 포괄적 입력변수로 구성한 데이터세트(Non-SA Set)와 구글의 자연어처리 API를 이용한 감성지수를 추가한 데이터세트(SA Set)로 구성하였다. 2개의 데이터세트에 4가지 모형을 학습시켜 총 8개의 예측모형을 개발한 결과는 다음과 같다. 첫째, SA Set에서는 로지스틱 모형 기반의 예측모형(LGST-SA)이 가장 높은 예측성과를 보였다. 둘째, Non-SA Set에서는 랜덤포레스트 모형 기반의 예측모형(RF-NSA)이 가장 높은 예측성과를 보였다. 셋째, LGST-SA의 예측성과는 RF-NSA보다 유의하게 높은 것으로 나타나, 감성지수의 예측기여도를 검증하였다. 추가적으로 해석가능한 기계학습 방법 중 하나인 SHAP을 통해 감성지수가 회계부정과 양(+)의 관계를 가지는 것을 사례적으로 확인하였다. 본 연구는 회계부정탐지모형의 개발에 있어 내부회계관리제도 운영보고서 감성분석의 유용성을 검증한 최초의 논문이며, 기계학습의 한계를 극복하기 위한 방안을 제시하였다는 점에서 의의가 있다.

This study tried to verify exploratorily whether the sentiment index of operation report of Internal Control over Financial Reporting (ICFR) has an additional contribution for developing accounting fraud detection models based on machine learning and to interpret the results. The sanction cases of Financial Supervisory Service (FSS) was used as a proxy for accounting fraud. Non-SA Set consists of comprehensive input variables, such as financial ratio and ICFR related variables (183) and SA Set is addition of the sentiment index of operation report of ICFR using Google's natural language process (NLP) API to Non-SA Set. The results of developing a total of 8 models by training 4 models on 2 each data sets are as follows. First, the logistic model based prediction model (LGST-SA) showed the highest predictive performance on SA Set. Second, the random forest model based prediction model (RF-NSA) showed the highest predictive performance on Non-SA Set. Third, the predictive performance of LGST-SA was significantly higher than that of RF-NSA, so the predictive contribution of the sentiment index was verified. Additionally, through SHAP, one of the interpretable machine learning (IML) methods, it was anecdotally confirmed that the sentiment index had a positive relationship on accounting fraud. This study has importance in that the first paper to verify the usefulness of the sentiment analysis of operation report of ICFR for developing an accounting fraud detection model and to suggest a way to overcome the limitations of the machine learning.