How to Scale Numerical Features

Numerical features often have vastly different ranges. For example, age may range from 18 to 90, while annual_income ranges from \$20,000 to \$500,000. Without scaling, models like Logistic Regression or SVM will assign disproportionate weight to large-magnitude columns.

This guide shows how to configure feature scalers in mlflow-experience.

---

Supported Scaling Strategies

Strategy	Scikit-Learn Class	Best Used For	Outlier Sensitivity
standard	StandardScaler	Features with a roughly normal (bell-curve) distribution. Centers at mean = 0 and std = 1.	Sensitive to extreme outliers.
minmax	MinMaxScaler	Bounded data or when you need features strictly within the [0, 1] range.	Highly Sensitive to outliers (compresses valid range).
robust	RobustScaler	Datasets with frequent outliers (e.g., salaries, transaction amounts). Centers using median and interquartile range (IQR).	Resistant to outliers.

---

Step 1: Configure Scaling in YAML

Define a scaling step in your preprocessing list:

preprocessing:
  - name: scale_features
    type: scaling
    params:
      columns:
        # Standard scaling for normally distributed features
        - name: age
          strategy: standard

        # Robust scaling for features with heavy tails or outliers
        - name: annual_income
          strategy: robust

        - name: total_spend
          strategy: robust

        # Min-Max scaling for scores with known boundaries
        - name: satisfaction_score
          strategy: minmax

        - name: credit_rating_percent
          strategy: minmax

---

How It Works

1. Grouped Execution:
2. mlflow-experience groups all columns by their chosen strategy and applies them using Scikit-Learn's ColumnTransformer.
3. Remainder Passthrough:
4. Any column not explicitly listed (e.g., one-hot encoded flags or IDs) is preserved untouched via remainder="passthrough".
5. No Data Leakage:
6. The scaler computes statistics (mean, std, median, min, max) only on the training split. During test evaluation or inference, it applies those learned parameters without recalculating on test data.

---

End-to-End Pipeline Order

For clean feature transformations, follow this logical sequence:

preprocessing:
  # 1. Impute missing numbers
  - name: handle_missing
    type: missing
    params:
      columns:
        - name: age
          strategy: median
        - name: annual_income
          strategy: median

  # 2. Scale numeric columns
  - name: scale_features
    type: scaling
    params:
      columns:
        - name: age
          strategy: standard
        - name: annual_income
          strategy: robust

  # 3. Encode categorical features
  - name: encode_categories
    type: encoding
    params:
      columns:
        - name: education_level
          strategy: one_hot

---

Step 2: Run the Pipeline

uv run python -m main --config configs/my_experiment.yaml --logging-level INFO

The model will train on appropriately scaled features, boosting convergence speed and prediction quality.