Imbalanced Semi-Supervised Learning via Label Refinement and Threshold Adjustment
Researchers have developed SEval, a novel framework designed to improve semi-supervised learning (SSL) performance on imbalanced datasets. Existing SSL methods often struggle with imbalanced data, leading to biased pseudo-labels that amplify majority class dominance. SEval addresses this by deriving theoretically optimal pseudo-labels and learning both pseudo-label refinement and threshold adjustment parameters from a class-balanced subset of the training data. Experiments show SEval outperforms current state-of-the-art methods in various imbalanced SSL scenarios, offering more accurate and reliable pseudo-labels. AI
IMPACT Enhances AI model performance on datasets with skewed class distributions, improving accuracy in real-world applications.