리샘플링과 클래스 가중치

You can fix skew by changing the data with resampling, or by changing the math with class weights. Both nudge attention to the rare class.

Oversampling copies or synthesizes more rare examples so the classes meet in the middle, giving the model more to learn from.

Undersampling drops some majority examples instead. It is fast and lean, but you risk throwing away useful signal.

SMOTE invents new minority points between real ones in feature space, rather than duplicating, so the model sees fresh variety.

Always rebalance the training set alone. Touching your test set fakes good scores and hides the real performance.

Instead of moving data, you can make each rare mistake hurt more. Higher class weights push the loss to respect the minority.

Many scikit-learn models accept class_weight. Set it to balanced and weights are computed inversely to class frequency.

Balanced weighting scales each class by the inverse of its frequency, so a rare label effectively counts many times more.

Class weights are cheap and keep your data intact, while resampling can capture richer patterns. Try weights first.

Aggressive oversampling can make a model memorize rare examples. Validate on untouched data to catch this overfitting early.

There is no universal winner. Choose a method, then judge it by minority recall, not by raw accuracy. 🎯

Pick the safest rebalancing habit.

Fight skew with resampling or class weights, rebalance training data only, and judge results on minority recall. ✅