Detecting exoplanet transits in Kepler light curves is challenging due to severe class imbalance and the highly localized morphology of transit dips. Standard oversampling approaches such as SMOTE generate synthetic minority samples through feature-space interpolation, which can distort or blur transit structures in phase-folded representations. We introduce a structure-preserving data augmentation method tailored to folded Kepler light curves. Our approach identifies transit dip regions in positive samples and applies controlled perturbations—phase jitter, constrained depth scaling, and realistic noise injection—while maintaining transit morphology consistency. Using 2048-bin folded signals and a 1D convolutional neural network trained under KIC-grouped splits to prevent target leakage, we compare the proposed method against a weighted baseline and SMOTE. Across 10 random seeds, the proposed soft augmentation configuration achieves higher and more stable AUPRC than SMOTE and provides consistent improvement over the weighted baseline. Ablation studies indicate that phase jitter is the primary contributor to performance gains, while overly aggressive depth perturbation can degrade results. These findings highlight the importance of domain-aware, structure-preserving augmentation for robust imbalanced classification of Kepler exoplanet candidates.
목차
Abstract 1. Introduction 2. Related Work 2.1. Deep Learning for Exoplanet Detection 2.2. Imbalanced Learning and Oversampling Techniques 2.3. Time-Series Data Augmentation 2.4. Physics-Informed and Structure-Aware Learning 2.5. Positioning of This Work 3. Method 3.1. Problem Formulation 3.2. Transit Structure Modeling 3.3. Structure-Preserving Augmentation Operator 3.4. Phase Jitter 3.5. Constrained Depth Scaling 3.6. Noise Injection 3.7. Targeted Minority Balancing 3.8. Learning Objective 3.9. Theoretical Motivation 4. Experiments 4.1. Dataset and Preprocessing 4.2. Grouped KIC-Based Splitting 4.4. Comparison Methods 4.5. Evaluation Metrics 4.6. Multi-Seed Stability Evaluation 4.7. Results Overview 4.8. Precision–Recall Curve Analysis 4.9. Ablation Study 5. Results and Discussion 5.1. Overall Performance Comparison 5.2. Why Does SMOTE Underperform? 5.3. Contribution of Augmentation Components 5.4. Precision–Recall Behavior 5.5. Stability Across Random Initializations 5.6. Physical Interpretation 5.7. Practical Implications 6. Conclusion References
조선대학교 기초과학연구원 [The Natural Science Research Institute of Chosun]
설립연도
2008
분야
자연과학>자연과학일반
소개
본 연구원은 기초과학을 진흥하기 위한 연구·교육 및 그 보급을 목적으로 한다. 이 목적을 달성하기 위하여 다음 각 호의 사업을 수행한다.
1. 기초과학 제 분야에 관한 조사와 연구
2. 기초과학에 관한 학술행사(학술대회, 학술세미나, 심포지엄, 초청강연회 등) 개최
3. 학문후속세대 및 일반인을 위한 기초과학 교육
4. 기관지『조선자연과학논문지』 발간
5. 『자연과학연구총서』, 『자연과학번역총서』 등 단행본 발간
6. 기타 본 연구원의 목적과 관련된 사업
간행물
간행물명
통합자연과학논문집(구 조선자연과학논문집) [Journal of Integrative Natural Science]