Skip to content

06 – Feature Scaling & Transformation

📘 RIDE User Manual – Panel 6: Feature Scaling & Transformation

📊 Purpose of the Panel

This panel allows users to normalize, standardize, or transform features for better performance in machine learning models. Proper scaling and transformation:

  • Makes algorithms converge faster.
  • Prevents features with larger magnitudes from dominating the model.
  • Handles skewness and non-normal distributions.

Recommended Reading

🧭 User Workflow

  1. Upload Dataset
    Choose from:

    • Initial DataFrame
    • After Missing Value Imputation
    • After Feature Encoding

  2. Select Scaling or Transformation Method
    Choose from 8 methods split into two categories:

    • Feature Scaling
    • Feature Transformation

  3. Choose Features to Scale
    Select one or more numeric columns.

  4. View Results

    • Preview scaled data
    • See summary stats before and after scaling
    • Compare original vs transformed distributions

  5. Download Scaled Data
    Download the transformed dataset for modeling.

💻 Features Breakdown

Feature Description
Source Selector Choose between raw, imputed, or encoded datasets.
Scaling Method Selector Includes both scaling and transformation strategies.
Numeric Column Detection Uses utility to auto-select numeric features.
Before/After Summary Side-by-side stats before and after scaling.
Distribution Comparison Interactive histogram to compare original vs scaled distributions.
Download Transformed Data Exports the result as a CSV.

🔧 Scaling & Transformation Methods

📏 Feature Scaling

Method Description Best For
Min-Max Scaling Rescales values to [0, 1] range. When bounded input is required (e.g., image pixel values).
Standardization (Z-score) Centers data with μ=0 and σ=1. When data needs normalization for gradient-based models.
Robust Scaler Uses IQR (Q3 - Q1), ignores outliers. When outliers are present and shouldn't dominate scaling.
Max AbsScaler Scales by max absolute value. When dealing with sparse data (e.g., TF-IDF).

🔄 Feature Transformation

Method Description Best For
Quantile Transformer Converts feature to a uniform distribution. When feature values are heavily skewed.
Log Transformer Applies log(1 + x). Right-skewed distributions (e.g., income, prices).
Power Transformer (Box-Cox) Normalizes data using λ parameter. Positive-only data with non-normal shape.
Power Transformer (Yeo-Johnson) Modified Box-Cox, supports negatives. Mixed-sign numeric data needing normalization.

🧠 Why This Panel Matters

  • Scaling ensures model convergence and fair weight distribution.
  • Transformation can reduce skewness and make data more Gaussian, which is preferred by many statistical and ML models.