Abstract
Class imbalance remains a central problem in machine learning because the rare class is often the event of scientific, clinical, financial, or operational interest. Yet the literature is fragmented across sampling, cost-sensitive learning, ensembles, representation learning, modern generative augmentation, and long-tail learning. This structured critical review examines these method families through the joint perspective of empirical class priors and data distributions. It distinguishes interventions that change the relative frequency of the classes from those that alter class-conditional distributions, meaning how feature values are distributed within each class, including minority support, local density, and class overlap. It also examines interventions that affect the information retained during training, the learning objective, the decision boundary, or the calibration of predicted probabilities, together with changes between the training and deployment distributions. The review draws on a documented pool of 470 bibliographic records, complemented by backward citation tracing and targeted searches to address identified gaps. Classical methods including the Synthetic Minority Over-sampling Technique (SMOTE), editing, Balanced Random Forest, EasyEnsemble, and random undersampling boosting (RUSBoost) are connected to focal and class-balanced losses, logit adjustment, decoupled training, generative adversarial network (GAN) /variational autoencoder (VAE)/diffusion synthesis, language-model tabular generation, graphs, federated learning, and domain shift. The synthesis identifies six recurrent failure mechanisms, namely, prior distortion, synthetic overlap, minority-noise amplification, unreliable high-dimensional neighborhoods, miscalibration, and evaluation leakage. The assumptions behind representative methods are presented, while an intervention map and conditional selection matrices translate the evidence into diagnostics, matched baselines, and leakage-free validation requirements. The resulting guidance relates method choice to the class priors observed during training and expected at deployment, the number and class-conditional distribution of minority examples, the degree of class overlap, the label structure, and whether the required output is a ranking, a class decision, or a calibrated probability.