Detecting and Classifying of Irregular Orders on Supply Chain Operation Using Isolation Forest = 공급망 운영상의 비정상 주문 탐지 및 분류를 위한 Isolation Forest 기반 연구

다국어 초록 (Multilingual Abstract)

The rapid digitalization of global supply chains—driven by the evolution of e-commerce, omnichannel logistics, and increasingly volatile consumer demand—has transformed the nature of operational data. Modern supply chains produce large volumes of continuous, high-frequency transactional information, capturing multiple dimensions of order behavior, shipment performance, and processing efficiency. Among these, essential operational features such as Unit quantity, Weight, Total Processing Time (TPT), Ship Late Day count, and Ship Ahead Day count serve as critical indicators of the flow, timing, and integrity of logistics execution. Although these data streams provide unprecedented visibility, they also introduce significant analytical challenges. Anomalous transactions—those that deviate sharply from expected behavior—can emerge from operational disruptions, forecasting inaccuracy, unexpected customer requirements, system errors, or data-entry inconsistencies. These anomalies have the potential to propagate through the supply chain, causing cascading effects such as bottlenecks, additional costs, and reduced service quality. Unlike traditional manufacturing and ERP-generated datasets, supply chain transactional data often lack explicit ground-truth labels identifying which records are erroneous or abnormal. Operational anomalies are rare, inconsistent, and context-dependent, making them difficult to capture through static thresholds or business rules. As a result, manual monitoring approaches are insufficient for timely and reliable anomaly detection, and many organizations remain vulnerable to hidden irregularities that are only discovered after they have already impacted performance. This motivates the development of automated anomaly detection systems that can learn latent behavioral patterns, adapt to dynamic conditions, and identify unusual transactions without relying on predefined labels. To address this challenge, the present study proposes a hybrid anomaly detection framework that integrates three complementary components: Isolation Forest (IF), XGBoost classification, and SHAP-based interpretability. Isolation Forest serves as the foundation of the framework, functioning as an unsupervised learning algorithm optimized for high-dimensional data with heterogeneous and skewed distributions—characteristics typical of supply chain operational datasets. After comprehensive preprocessing steps including normalization, encoding of structural variables, and feature selection, the Isolation Forest model was trained with a contamination rate of 0.1, aligning with the anomaly proportion observed during exploratory analysis. The model produced anomaly scores by isolating observations through recursive random partitioning, allowing transactions with shorter isolation path lengths to be identified as anomalous. Visualization of the anomaly score distribution revealed a strong structural separation between normal and abnormal transactions, with a distinct right-tailed distribution that highlighted extreme deviations in processing time, shipping behavior, and physical order characteristics. These results indicate that Isolation Forest effectively captures the core irregularities embedded in supply chain data, detecting anomalies that are both operationally meaningful and statistically significant. To further refine the detection boundary and enhance interpretability, a semi-supervised learning layer using XGBoost was applied. The anomaly labels generated by Isolation Forest acted as pseudo-targets, enabling XGBoost to learn complex relationships and nonlinear interactions among variables that contribute to unusual behavior. This additional step produced smoother anomaly likelihoods compared to binary isolation outcomes, improving the stability and interpretability of the detection process. The high computational efficiency of XGBoost also ensured that the framework remained scalable and suitable for large, rapidly growing operational datasets. Interpretability was a central objective of the proposed framework, given the importance of transparency in logistics decision-making. To achieve this, SHAP (SHapley Additive exPlanations) values were employed to quantify the contribution of each feature to the anomaly predictions generated by XGBoost. The SHAP analysis revealed that TPT, Ship Late Day count, Ship Ahead Day count, Weight, and Unit quantity were consistently the most influential factors. These findings align with practical supply chain understanding: prolonged processing times often indicate workflow congestion or system inefficiencies, while extreme deviations in shipping timing may reflect forecasting errors, inventory shortages, or transportation breakdowns. SHAP provided both global insights into the overall model behavior and local explanations for individual anomalous transactions, empowering supply chain managers to diagnose the root causes of operational irregularities. A comparative analysis was also conducted across three unsupervised anomaly detection algorithms—Isolation Forest, One-Class SVM (OC-SVM), and Local Outlier Factor (LOF)—to evaluate agreement and reliability across methodological perspectives. The overlap analysis indicated that 214 transactions were jointly detected as anomalies by all three models, demonstrating a strong core consensus among techniques with fundamentally different detection mechanisms. While LOF detected the largest number of unique anomalies due to its focus on local density variations, and OC-SVM flagged a moderate number influenced by boundary separation in high-dimensional space, Isolation Forest identified the fewest unique anomalies. However, the anomalies isolated by IF were also the most structurally pronounced, demonstrating that the model prioritizes globally meaningful deviations rather than noise-driven fluctuations. This selective precision, combined with its strong alignment with the multi-model intersection, reinforces Isolation Forest as the most robust and reliable anomaly detection approach in the context of supply chain operations. The results of this research confirm that the proposed hybrid framework is both effective and operationally valuable. Isolation Forest provides a strong unsupervised foundation, XGBoost enhances detection boundaries through predictive learning, and SHAP ensures interpretability critical for managerial decision-making. Together, the components form a coherent, scalable, and transparent pipeline suitable for real-world logistics environments that require timely, explainable, and actionable insights.

국문 초록 (Abstract)

전자상거래의 발전, 옴니채널 물류의 확산, 그리고 변동성이 점점 증가 하는 소비자 수요로 인해 글로벌 공급망은 빠르게 디지털화되고 있다. 이러한 변화로 인해 현대 공급망은 주문 행태, 운송 성과, 처리 효율성 과 관련된 다양한 차원의 고빈도 거래 데이터를 대량으로 생성하고 있 다. 특히 단위 수량, 중량, 총 처리 시간, 배송 지연 일수, 배송 조기 일 수와 같은 운영 지표는 물류 실행의 흐름, 시점, 그리고 데이터 무결성 을 파악하는 데 있어 핵심적인 역할을 한다. 그러나 이러한 데이터의 증 가는 동시에 분석상의 복잡성과 도전 과제를 야기한다. 운영 중단, 수요 예측 오류, 예기치 못한 고객 요구, 시스템 오류, 또는 데이터 입력 오류 로 인해 정상적인 패턴에서 크게 벗어난 이상 거래가 발생할 수 있으며, 이러한 이상 거래는 병목 현상, 추가 비용 발생, 서비스 품질 저하 등 공급망 전반에 연쇄적인 부정적 영향을 미칠 수 있다. 전통적인 제조 및 ERP 기반 데이터와 달리, 공급망 거래 데이터는 이 상 여부에 대한 명확한 정답 레이블이 존재하지 않는 경우가 많다. 운영 상의 이상 현상은 발생 빈도가 낮고, 상황 의존적이며, 불규칙하게 나타 나기 때문에 고정된 임계값이나 규칙 기반 방식으로는 효과적으로 포착 하기 어렵다. 그 결과, 수작업 중심의 모니터링 방식은 적시성과 신뢰성 이 부족하며, 많은 조직이 성과에 이미 영향을 미친 이후에야 숨겨진 이상 현상을 인지하는 한계를 지닌다. 이러한 배경에서, 사전 정의된 레이 블에 의존하지 않고 잠재적인 패턴을 학습하며, 동적인 환경 변화에 적 응할 수 있는 자동화된 이상 탐지 시스템의 필요성이 제기된다. 본 연구는 이러한 문제를 해결하기 위해 Isolation Forest(IF), XGBoost 분류 모델, 그리고 SHAP(SHapley Additive exPlanations) 기반 해석 기법 을 결합한 하이브리드 이상 탐지 프레임워크를 제안한다. Isolation Forest는 고차원적이고 이질적이며 분포가 왜곡된 공급망 데이터의 특성 에 적합한 비지도 학습 기반 알고리즘으로, 본 프레임워크의 핵심 탐지 단계로 활용된다. 데이터 정규화, 구조적 변수 인코딩, 특성 선택을 포함 한 전처리 과정을 거친 후, 오염 비율을 0.1로 설정하여 모델을 학습하 였으며, 이는 탐색적 분석을 통해 관찰된 이상 거래 비율을 반영한 것이 다. Isolation Forest는 무작위 분할을 통해 관측치를 격리시키며, 더 적은 분할로 격리되는 거래일수록 이상 거래로 간주하여 이상 점수를 산출한 다.이상 점수 분포를 시각화한 결과, 정상 거래와 이상 거래 간에 구조적 으로 뚜렷한 분리가 나타났으며, 특히 처리 시간, 배송 행태, 주문의 물 리적 특성에서 극단적인 편차를 보이는 거래들이 오른쪽 꼬리를 형성하 는 분포를 확인할 수 있었다. 이는 Isolation Forest가 공급망 데이터에 내재된 핵심적인 이상 패턴을 효과적으로 포착하고 있음을 의미한다. 탐지 결과의 안정성과 해석 가능성을 향상시키기 위해, 본 연구는 Isolation Forest로부터 생성된 이상 레이블을 의사 레이블로 활용하여 XGBoost 기반의 준지도 학습 단계를 추가하 였다. 이를 통해 변수 간의 비선형적 관계와 복합적인 상호작용을 학습 할 수 있었으며, 단순한 이진 분류 결과보다 부드러운 이상 가능성 분포 를 도출함으로써 탐지 경계의 안정성을 개선하였다. 또한 XGBoost의 높 은 계산 효율성은 대규모 거래 데이터 환경에서도 본 프레임워크의 확장 가능성을 보장한다. 공급망 의사결정에서 투명성과 설명 가능성의 중요성을 고려하여, 본 연구는 SHAP기법을 적용하여 이상 분류 결과에 대한 해석을 수행하였다. SHAP 분석 결과, 총 처리 시간, 배송 지연 일수, 배송 조기 일수, 중 량, 단위 수량이 이상 거래 판별에 가장 큰 영향을 미치는 변수로 나타 났다. 이는 처리 시간이 길어질수록 운영 병목이나 시스템 비효율이 발 생할 가능성이 높아지고, 배송 시점의 극단적인 편차가 수요 예측 오류, 재고 부족, 또는 운송 장애를 반영할 수 있다는 실무적 이해와도 일치한 다. SHAP는 모델 전체 수준에서의 전반적인 영향 요인뿐만 아니라, 개별 거래 단위에서의 국지적 설명을 제공함으로써 이상 거래의 원인을 진단 할 수 있는 실질적인 도구로 활용된다. 또한 본 연구는 Isolation Forest, One-Class SVM(OC-SVM), Local Outlier Factor(LOF) 세 가지 비지도 이상 탐지 기법 간의 비교 분석을 수행하였다. 본 연구에서는 비지도 머신러닝 기반 이상 탐지 기법인 Isolation Forest, One-Class SVM, Local Outlier Factor를 적용하여 비정 상 거래를 탐지하였다. 각 기법은 데이터 분포 가정, 이상치 정의 방식 및 탐지 특성에서 차이를 가지므로, 동일한 거래 데이터셋에 적용하여 탐지 성능과 실무 적용 적합성을 비교·분석하였다. 이를 통해 거래 이 상 탐지에 있어 비지도 학습 기법의 효용성을 객관적으로 검증하고자 한 다.각 기법은 데이터 분포 가정, 이상치 정의 방식 및 탐지 특성에서 차 이를 가지므로, 동일한 거래 데이터셋에 적용하여 탐지 성능과 실무 적 용 적합성을 비교·분석하였다. 이를 통해 거래 이상 탐지에 있어 비지 도 학습 기법의 효용성을 객관적으로 검증하고자 한다분석 결과, 서로 다른 탐지 메커니즘을 가진 세 모델이 공통적으로 이상으로 분류한 거래 는 총 214건으로, 핵심적인 이상 거래 집합에 대한 높은 합의를 보였다. LOF는 지역 밀도 변화에 민감하여 가장 많은 고유 이상 거래를 탐지한 반면, OC-SVM은 고차원 공간에서의 경계 설정에 따라 중간 수준의 탐 지 결과를 보였다. Isolation Forest는 고유 이상 거래 수는 가장 적었으 나, 구조적으로 가장 뚜렷하고 의미 있는 이상 거래를 우선적으로 탐지 하였다. 이러한 결과는 Isolation Forest가 잡음에 의한 변동보다는 전역 적으로 중요한 편차를 효과적으로 식별함을 시사한다. 종합적으로 본 연구의 결과는 제안된 하이브리드 프레임워크가 효과적 일 뿐만 아니라, 실무적으로도 높은 활용 가치를 지님을 보여준다. Isolation Forest는 강력한 비지도 탐지 기반을 제공하고, XGBoost는 탐지 경계를 정교화하며, SHAP는 물류 의사결정에 필수적인 해석 가능성을 확보한다. 이 세 구성 요소는 상호 보완적으로 결합되어, 실제 물류 환 경에서 요구되는 신속성, 확장성, 그리고 설명 가능성을 동시에 충족하 는 이상 거래 분석 파이프라인을 형성한다.

목차 (Table of Contents)

• I. INTRODUCTION ·····························································································1
• II. LITERATURE REVIEW ·············································································12
• 1. Anomalies in Supply Chain Operations ··················································12
• 2. Traditional Rule-Based and Statistical Approaches ·····························14
• 3. Machine Learning Methods for Anomaly Detection ····························15
• 4. Concept and Mechanism of Isolation Forest ·········································16
• 5. Comparison with One-Class SVM and Local Outlier Factor ······················18
• III. DATA & METHODOLOGY ·····································································20
• 1. General information about dataset ·························································20
• 2. Isolation Forest ····························································································23
• 3. Classification and interpretability using XGBoost and SHAP ·····················26
• 4. Performance metrics ··················································································28
• 5. Model Integration Framework ···································································30
• IV. RESULTS & ANALYSIS ···········································································34
• 1. Result of Isolation Forest ··········································································34
• 2. Anomaly Score Distribution ·······································································36
• 3. Result of XGBoost and SHAP ··································································38
• 4. Resul of Three Methods (IF, OC-SVM, and LOF) ·······························40
• 5. Boxplot Comparison of Anomaly Scores Across IF, OC-SVM, and LOF ····41
• 6. Discussion and interpretation of differences ········································42
• V. CONCLUSION ·······························································································46
• 1. Summary of findings ··················································································46
• 2. Future works ································································································49
• REFERENCES ···································································································51
• 논문요약 ············································································································55