Imbalanced Data — Resampling and SMOTE

Class imbalance, where one outcome vastly outnumbers another, is the silent saboteur of classification, fooling both models and the metrics used to judge them. Handling imbalance exists as a discipline because most classifiers and the default accuracy metric implicitly assume balanced classes, so on imbalanced data a model can achieve high accuracy by simply predicting the majority class while completely failing at the rare class that usually matters most — the fraud, the disease, the defect. Without deliberate handling, the minority class is drowned out, and the model that looks excellent on accuracy is useless for its actual purpose. The remedies — resampling the data, synthesising minority examples with SMOTE, and reweighting the loss — each rebalance the learning differently, and combined with appropriate metrics they are what make models genuinely capable of detecting the rare events they were built to find.