<한양대학교 양승지 박사님 초청 강연 안내>

1. 연사: 한양대학교 양승지 박사님

2. 주제: Imbalanced Classification: From Geometric Heuristics to Statistical Approaches

3. 일시: 2026년 10월 8일(목) 16:00

4. 장소: 초청 강의실(자1-124)

5. 초록:

Class imbalance, where one class is substantially underrepresented relative to the other, is a common challenge in real-world binary classification, such as rare disease diagnosis and fraud detection. While Bayes decision theory provides a clean foundation for understanding optimal classifiers, standard methods break down under severe imbalance: the decision boundary is pushed away from the minority class, and the Bayes-optimal classifier is poorly approximated even with large samples. Oversampling — generating synthetic minority instances — is one of the dominant approaches, yet most existing methods rely on geometric heuristics and largely ignore the distributional structure of the minority class. In the first part of this talk, I will introduce the problem of class imbalance and survey the landscape of synthetic oversampling methods, from SMOTE to its many variants. The second part will present our efforts to reformulate oversampling as a problem of conditional density estimation, introducing a Gaussian mixture model-based framework with statistically grounded overlap quantification and an ongoing extension toward robust estimation and total variation distance.