Most real classification problems are imbalanced. Fraud is 0.1%, disease positive is 2%, churn is 5%—the minority class is what we actually want to catch. Training on raw imbalanced data causes models to achieve high accuracy just by predicting all majority class.
Today we explore why imbalance is a trap, and how to handle it via resampling (over/undersampling), SMOTE, class weights, and metric selection. This is a frequently-tested MLS-C01 domain.
With 2% positive data, predicting "all negative" yields 98% accuracy—yet catches zero positives we care about. This is the accuracy paradox.