희소 클래스가 무시되는 이유

When one label hugely outnumbers another, your data is imbalanced. Think 9,500 normal emails versus 500 spam ones.

A model wants high accuracy fast. The easy win is to always predict the majority class and quietly ignore the rare one. 😬

With 5% spam, a model that labels everything ham scores 95% accuracy yet catches zero spam. That number is hollow.

Standard training minimizes total error. Since rare-class mistakes are few, the loss barely moves when the model ignores them.

We call the big group the majority class and the small one the minority class. NLP often cares most about the minority.

Before modeling, count your labels. One line reveals how lopsided the class distribution really is.

A missed spam, fraud, or abuse message can be costly. In NLP the rare class is usually the one you built the model to catch.

With few minority points, the decision boundary drifts toward the crowd, swallowing the rare region almost entirely.

On skewed data, accuracy hides failure. You need metrics that spotlight the rare class, like recall on the minority.

A quick imbalance ratio tells you the severity. Ten-to-one is mild; a thousand-to-one needs serious care.

Naming the problem is half the battle. Once you see the skew, you can choose resampling or weighting to fight it.

Let us test the core idea behind imbalance.

Imbalanced data lets models ignore the rare class while scoring high accuracy. Spot the skew first, then fix it. ✅