숫자 크기 조정과 정규화

Age ranges 0 to 90 while income ranges to millions. Without scaling, the bigger numbers dominate and quietly drown out the smaller ones. ⚖️

Models that measure distance, like k-NN and k-means, are very sensitive to scale. Unscaled features hand all the power to the largest column.

Standardization rescales a column to mean 0 and standard deviation 1. Values become how many standard deviations they sit from the average.

The StandardScaler applies standardization for you. Fit it on your data, then transform any column into z-scores.

Normalization squeezes values into a fixed range, usually 0 to 1. The smallest value maps to 0 and the largest to 1.

Reach for MinMaxScaler to rescale into 0 to 1. It keeps the shape of your data while bounding the range.

Use standardization when data is roughly bell-shaped or has outliers. Pick normalization when you need a strict 0 to 1 bound.

A single huge value can squash everything else near zero under MinMax. When outliers dominate, RobustScaler resists them better.

Fit the scaler on training data alone, then transform the test set. Fitting on everything leaks test information into your model.

Bundle the scaler with your model in a Pipeline. It then fits only on the training fold during cross-validation, blocking leakage for free.

Tree-based models like random forests split on thresholds, so scaling rarely changes results. Save scaling for distance and linear methods.

You need every feature rescaled to a strict 0-to-1 range. Which scaler fits?

You put features on a fair footing: StandardScaler for z-scores, MinMaxScaler for 0 to 1, RobustScaler for outliers. Always fit on train only. 🎉