Active Learning Strategies for the Class Imbalance Problem

초록

Recent advancements in data technology have greatly expanded data availability; however, extracting valuable information remains challenging due to high annotation costs and severe class imbalance. Active learning (AL) addresses this by minimizing annotation costs while maximizing model performance through selectively querying informative samples. While class imbalance often deteriorates model performance, it is prevalent in real-world scenarios. This study provides a comprehensive survey of the class-imbalance problem within active learning, analyzing regular articles published from 2018 to May 2025. We specifically posit that active learning can mitigate the imbalance problem as effectively as traditional approaches. Based on seminal literature, we propose a five-category taxonomy for imbalance mitigation: Data-level, Probability-based, Weight-adjustment, Loss-based, and Optimization approaches. Furthermore, we explore application-specific methods, evaluate key metrics, and highlight open challenges. To the best of our knowledge, this is the first topical review focusing on active learning for class-imbalance, offering valuable insights to the machine learning community. © 2013 IEEE.

키워드