정규화 강도 조정하기

With thousands of word features, a model can memorize quirks of the training set. Regularization keeps it from chasing that noise. 🧯

Regularization adds a penalty for large coefficients. The model must justify every big weight, so it stays simpler and generalizes better.

In scikit-learn you control strength with C. It is the inverse of regularization, so smaller C means a stronger penalty.

A tiny C shrinks weights hard toward zero. The model becomes very simple, which can underfit and miss real signal.

A big C weakens the penalty and lets weights grow. The model fits training data closely but risks overfitting.

The default L2 penalty gently shrinks all weights. It is a safe starting point for most text classification problems.

Switching to the L1 penalty zeroes out many coefficients, performing feature selection and giving you a sparser, simpler model.

Do not guess C by hand. Use cross-validation to test several values and pick the one that scores best on held-out folds.

GridSearchCV tries each C, runs cross-validation, and reports the winner. Sweep C across powers of ten for a fast first pass.

After fitting, best_params_ shows the C that won. The matching model is ready for honest evaluation on your test set.

The goal is balance: enough freedom to learn, enough penalty to generalize. That sweet spot gives the best score on unseen text.

What happens as you lower the C value?

Regularization fights overfitting, and C sets its strength inversely. Tune C with cross-validation to find the balanced sweet spot. ✅