Class-Difficulty Aware Hybrid Active Learning

Abstract

Active learning (AL) aims to select the most valuable samples from an unlabeled dataset for annotation, maximizing model performance under a limited sample budget. Identifying these samples becomes even more challenging in high-dimensional data scenarios. On the other hand, while samples from difficult classes are more informative than those from simple classes, most existing AL methods do not account for the impact of class difficulty on sample uncertainty. To address these challenges, we propose a two-stage hybrid active learning method CDHAL. In the first stage, we use class-difficulty coefficients to weight the uncertainty scores of samples and select highly uncertain samples to form a candidate pool. In the second stage, we use a core-set method based on probabilistic distance between samples to select samples with high informativeness and low redundancy from the candidate pool for annotation. Experimental results demonstrate that our CDHAL outperforms state-of-the-art methods on various benchmark datasets.