Class Imbalance

Class imbalance describes a dataset where target categories appear in very different proportions. It is common in fraud, medical diagnosis, security incidents, equipment failure, and other tasks where the important event is rare. A model trained to optimize overall accuracy may ignore the minority class and still receive a high score. Responses include resampling, class-weighted losses, anomaly-detection methods, threshold adjustment, targeted data collection, and evaluation with precision, recall, F1, or precision–recall curves. The correct approach depends on error costs and real prevalence. Artificially balanced training data can help learning, but predicted probabilities may need recalibration before deployment in the original population.