Multi-Label Classification

Multi-label classification predicts a set of labels for each example, useful when images, documents, or clinical records can have several valid categories at once. Common approaches include independent binary classifiers, classifier chains, and neural models that capture label relationships with attention, graphs, or hierarchies; large label spaces also use retrieval and generative methods. Work addresses long-tailed labels, incomplete or noisy annotations, and changing domains through reweighting, augmentation, semi-supervised learning, and test-time adaptation. A central tension is that better aggregate scores do not guarantee reliable rare-label predictions: thresholds, calibration, uncertainty, and evaluation under distribution shift remain important challenges.