Handling Imbalanced Data

Imbalanced Classification

When classes have unequal representation, standard metrics are misleading.

Resampling Techniques

Oversampling increases minority class: RandomOverSampler. Undersampling reduces majority: RandomUnderSampler.

SMOTE creates synthetic minority samples: SMOTE(). It interpolates between minority samples.

Class Weights

class_weight='balanced' adjusts weights inversely proportional to class frequencies. This penalizes minority errors more.

Works with most classifiers. Simple to implement.

Evaluation

Use AUC, F1, precision-recall curve, not accuracy. Accuracy can be high with trivial models on imbalanced data.

Confusion matrix reveals prediction patterns. Classification report shows per-class metrics.

Key Takeaways

1. Resampling balances class representation
2. Class weights adjust algorithm behavior
3. Choose metrics appropriate for imbalanced data