Macro-AUC-Driven Active Learning Strategy for Multi-Label Classification Enhancement
Minghao Li, Junjie Qiu, Weishi Shi · 2024
Data annotation for multi-label tasks is often significantly more expensive than traditional machine learning tasks (such as multi-class classification) due to the high label cardinality. In this paper, we propose a new active learning (AL) method designed to reduce the data annotation cost for multi-label tasks. Unlike other loss-agnostic multi-label AL methods, our approach is specifically designed for Macro-AUC optimization, based on Multi-label Ranking loss. This specialization elucidates the nuanced interplay between multi-label classification and Macro-AUC optimization. We refine the dataset by selecting the most diverse samples from the unlabeled data, ensuring relevance and diversity in the training set. Then an uncertainty-based selection strategy is adopted, focusing on the most uncertain samples under the most uncertain label. We conduct an intuitive analysis to justify the effectiveness of our proposed AL strategy.