ปรับขนาดและเลือกคุณลักษณะ

Text models can explode to tens of thousands of features. Many add only noise, so trimming the list often improves both speed and accuracy.

Some models compare feature magnitudes directly. If one feature ranges 0 to 1000, it can drown out the rest unless you scale them first.

StandardScaler shifts each feature to zero mean and unit variance. Now every numeric feature speaks on the same scale.

Centering a sparse TF-IDF matrix fills it with nonzeros and wastes memory. Use MaxAbsScaler, which scales without destroying sparsity.

Features that barely vary tell the model nothing. A VarianceThreshold filter removes near-constant columns in one quick pass.

SelectKBest keeps the top features by a scoring test like chi-squared. You ask for the best k and it drops the rest.

An L1-penalized model pushes useless weights to zero on its own. This built-in selection is often called embedded feature selection.

Always fit scalers and selectors on training data alone, then apply them to test data. Mixing them causes leakage and rosy fake scores.

Putting scaling and selection inside a Pipeline runs them in the right order every time. It also blocks accidental leakage during validation.

A leaner feature set trains faster, needs less memory, and is easier to explain. Smaller is often better once noise is gone. ⚡

Try a few values of k and compare validation scores. Let the numbers, not a hunch, decide how many features to keep.

Which scaler keeps a sparse matrix sparse?

Scale numeric features and select the useful ones with SelectKBest or L1. Use MaxAbsScaler for sparse data, and fit everything inside a pipeline. ✅