Feature Engineering

Why Scale Features?

• Gradient descent converges faster
• Distance metrics work correctly
• Regularization applies equally

Min-Max Normalization

• x' = (x - min) / (max - min)
• Scales to [0, 1]
• Use for bounded inputs

Standardization

• z = (x - mean) / std
• Mean=0, Std=1
• Use for different units

One-Hot Encoding

• Binary column per category
• [0, 1, 0] for "blue"
• No false ordering

Missing Values

• Drop rows (loses data)
• Impute with mean/mode
• Add "was missing" indicator

Feature Creation

• Polynomial: x², x₁×x₂
• Log transform for skewed data
• Domain-specific combinations

Feature Selection

• Filter: correlation, mutual info
• Wrapper: try feature subsets
• Embedded: L1 regularization