Decision Tree - Missing And Unknown Values

Handling Missing Values

Decision trees can handle missing values in several ways:

1. Surrogate splits: Use alternative features that correlate with the primary feature
2. Separate category: Treat missing values as a separate category
3. Most common value: Impute with mode (classification) or mean (regression)

Handling Unknown Values (Non-NA)

For categorical features with values not seen during training:

• Assign to the majority class of the parent node
• Use a default path determined during training
• Treat as a separate “unknown” category