Scale Numeric Features (Auto-Select Method)

Fits a scaler on the training set and applies the same parameters to both train and test sets, preventing data leakage.

Usage

scale_data(split_result, method = "auto", cols = NULL, verbose = TRUE)

Arguments

split_result

The list returned by split_data().

method

Character. One of "auto" (default), "minmax", "zscore", or "robust". Use "auto" to let the function choose.

cols

Character vector. Names of numeric columns to scale. NULL (default) scales all numeric columns.

verbose

Logical (default TRUE).

Value

An invisible list with elements:

train_scaled

Scaled training data.frame.

test_scaled

Scaled test data.frame.

params

Per-column scaling parameters (fitted on train only).

method

The scaling method used.

method_reason

Why the method was selected (auto mode only).

outlier_ratio

Named numeric vector of outlier ratios per column.

cols

Character vector of scaled column names.

Details

When method = "auto" (default), the scaling method is chosen automatically based on per-column outlier detection (IQR method) and a normality test (Shapiro-Wilk):

• robust – any outlier found in any column (IQR fence)

• zscore – no outliers + majority of columns approximately normal

• minmax – no outliers + majority of columns not normal

Methods

minmax

Scales each feature to \([0, 1]\).

zscore

Standardises to zero mean and unit variance.

robust

Uses median and IQR, robust to outliers.

Examples

data(airquality)
clean <- airquality[complete.cases(airquality), ]
sp    <- split_data(clean, verbose = FALSE)
sc    <- scale_data(sp)
#> 
#> ============================================================
#>   STEP 7/7 : Feature Scaling  [ROBUST]
#> ============================================================
#> 
#>   Auto-selected : outlier detected in 2 column(s) [Ozone, Wind] -> robust
#>   Columns : 6  (Ozone, Solar.R, Wind, ...)
#>   Fitted on TRAIN only  |  Columns with outliers: 2 / 6
#> 
#> ------------------------------------------------------------
#>   Outlier ratio per column  (IQR)
#> ------------------------------------------------------------
#>  column outlier_pct
#>    Wind       3.41%
#>   Ozone       1.14%
#> 
#>   [OK] Outliers found in [Ozone, Wind] -> robust scaling applied
#> 
sc$method
#> [1] "robust"
sc$method_reason
#> [1] "outlier detected in 2 column(s) [Ozone, Wind] -> robust"