Long-Tailed Classification
Overview
Long-tailed classification addresses the tendency of models trained on imbalanced data to favor common classes and miss rare but important ones. Methods include class reweighting and logit adjustment, balanced or progressive sampling, classifier retraining, contrastive and prototype-based representation learning, and synthetic data augmentation. Research spans image recognition, medical multi-label diagnosis, text, and semi-supervised learning, with evaluation increasingly examining tail-class recall, calibration, subgroup fairness, and performance under domain shift. A central tension is improving rare-class recognition without sacrificing overall accuracy or overfitting scarce examples; no single approach works best across all data and settings.
Latest papers
55 papers6 papers in the last four weeks, level with the four weeks before. 0.0% of all new papers.