사용자–환경 상호작용 데이터에 대한 쾌적구간 기반 해석가능한 군집화 방법 = A Comfort-Zone-Based Interpretable Clustering Method from User–Environment Interaction Data

다국어 초록 (Multilingual Abstract)

Modern users continuously interact with various physical environments such as residential HVAC systems, office lighting, and vehicle climate control, actively adjusting devices to maintain personal comfort. These user-environment interactions significantly impact quality of life, health, work productivity, and energy efficiency. With the advancement of Internet of Things (IoT) technology and widespread adoption of smart devices, large-scale collection of user behavior logs has become possible, enabling research on personalized services, energy management, and smart environment design. However, user-environment interaction data poses inherent analytical challenges including the absence of explicit labels for subjective states like comfort or satisfaction, individual differences and temporal variability in preferences, ambiguous behavioral signals where device operations may reflect discomfort, habitual actions, or exploratory behavior, and domain heterogeneity across different systems with varying sensor types, measurement units, and temporal resolutions. Existing approaches are limited by either requiring continuous collection of satisfaction surveys for supervised learning methods or, in unsupervised approaches, suffering from domain-specific designs, failure to separate signal from noise in adjustment events, and lack of interpretability in discovered user groups or patterns. To address these limitations, this thesis proposes the CUE (Comfort-zone-driven User clustering for Explainable preference analysis) framework, which infers users' latent preference structures from unlabeled behavioral logs and represents them in an interpretable and generalizable form. The framework integrates three core mechanisms: domain-generalized pattern discovery through Context-Action (CA) structure that unifies heterogeneous log formats from HVAC, lighting, and vehicle systems; reliability-based pre-clustering that filters sporadic and exploratory operations through Adjustment Frequency Index (AFI) and Preference Consistency Index (PCI), then extracts personalized comfort zones representing environmental ranges maintained without user adjustment; and interpretable feature-based clustering that transforms comfort zone boundaries into two-dimensional feature vectors with physical meaning to derive final user clusters. The framework operates through six sequential stages: data structuralization converting domain-specific logs to CA structure, metric computation calculating AFI and PCI for each environmental value, threshold determination establishing criteria for meaningful adjustments based on distribution characteristics, comfort zone extraction determining personalized boundaries exceeding thresholds, feature extraction representing zones as 2D vectors, and clustering grouping users with similar environmental preferences. Experimental validation across cooling, heating, and lighting domains demonstrates that personalized comfort zones achieve 133% improvement in Dissatisfaction Activation Rate (DAR) and 54% improvement in F1 score compared to fixed-range baselines, accurately reflecting actual user behavior. The comfort zone-based feature representation shows 27% average improvement in Silhouette Coefficient (SC) and rule stability compared to existing feature representations, simultaneously enhancing cluster quality and interpretability. The framework operates consistently across all three domains, proving domain generalization capability. Notably, comfort zone-based features form clusters directly corresponding to physical variables (e.g., "low-temperature preference group," "high-temperature preference group"), maintain consistent structure across various visualization techniques including t-SNE and UMAP, and demonstrate robust performance across varying cluster numbers, confirming applicability in real system operation environments. This research provides a practical methodology for establishing personalized environmental control strategies using only implicit behavioral signals without explicit preference labels in IoT environments, expected to contribute to user experience improvement and energy efficiency enhancement in various environmental control applications including smart homes, building automation, and vehicle climate systems.

목차 (Table of Contents)

• I. 서 론 1
• II. 배 경 7
• 2.1. 사용자-환경 상호작용의 이해 7
• 2.1.1. 상호작용 데이터의 정의와 특성 7
• 2.1.2. 다양한 도메인에서의 상호작용 7
• 2.2. 전통적인 환경 쾌적성 평가 9
• 2.3. 특징 표현과 추출 12
• 2.3.1. 특징 공학의 개념 12
• 2.3.2. 대표적인 특징 공학 방법 13
• 2.3.3. 해석 가능한 특징 표현 15
• 2.4. 임계값 기반 데이터 분할 16
• 2.5. 개인화 모델링 방법론 19
• 2.5.1. 사용자 선호 모델링의 기초 19
• 2.5.2. 지도학습 기반 접근 20
• 2.5.3. 비지도 학습 기반 접근 21
• 2.6. 사용자 유형화에 대한 기존 연구들과의 비교 22
• III. 문제 정의 25
• IV. 제안하는 CUE 프레임워크 32
• 4.1. 도메인 일반화를 위한 데이터 구조화 34
• 4.2. 신뢰성 검증을 위한 지표 설계 37
• 4.2.1. 설계 원칙 및 요구사항 37
• 4.2.2. 조정 빈도수 지수 (AFI) 38
• 4.2.3. 선호도 일관성 지수 (PCI) 39
• 4.3. 분포 기반 지표 임계값 결정 43
• 4.4. 개인화된 쾌적구간의 동적 산출 44
• 4.4.1. 쾌적구간의 개념과 구조 44
• 4.4.2. 쾌적구간의 수학적 정의 46
• 4.5. 쾌적구간 기반 사용자 사전 군집화 48
• 4.6. 쾌적구간 기반 특징 추출 50
• 4.7. 유사도 기반 사용자 군집화 50
• V. 실험 결과 및 분석 52
• 5.1. 실험 개요 52
• 5.1.1. 실험 목표 및 연구 질문 52
• 5.1.2. 실험 환경 및 구현 52
• 5.1.3. 실험 데이터셋 54
• 5.2. AFI 및 PCI의 임계값 산출 56
• 5.3. 개인화된 쾌적구간의 행동 반영도 평가 58
• 5.3.1. 연구질문 1에 대한 실험 설계 58
• 5.3.2. 평균 행동 반영도 분석 63
• 5.3.3. 시간에 따른 행동 반영도 추세 분석 67
• 5.4. 쾌적구간 기반 특징의 군집 해석 가능성 평가 69
• 5.4.1. 연구질문 2에 대한 실험 설계 69
• 5.4.2. 평균 군집 품질 및 해석 가능성 분석 74
• 5.4.3. 군집화 결과에 대한 시각적 분석 79
• 5.4.4. 시간에 따른 군집 품질 안정성 분석 84
• 5.5. 도메인 일반화 성능 종합 분석 86
• VI. 결 론 88
• VII. 향후 연구 91
• 참고 문헌 92
• 영문 초록 98