Binary Classification

Binary classification studies how to learn decision rules that separate two classes from labeled data, and how to make their predictions useful when labels are scarce, noisy, or unevenly distributed. Common approaches include linear models and support vector machines, trees and boosting, neural networks, and positive-unlabeled learning, alongside threshold adjustment, reweighting, and abstention. Theoretical work examines generalization, calibration, privacy, and robustness, while empirical studies use settings ranging from tabular and biomedical data to text, images, and security tasks. A central tension is that accuracy or AUC alone may not predict performance at operational thresholds, where minority-class recall, false alarms, calibration, and distribution shift matter.