Train/Test Split

Never evaluate on training data! Splitting data properly is essential for honest evaluation.

Why Split Data?

The Standard Split

Shuffling and Stratification

Key Takeaways

• Always split data BEFORE any preprocessing that uses y
• Use 80/20 or 70/30 split for most cases
• Shuffle to avoid ordering bias
• Stratify for classification with imbalanced classes
• Set random_state for reproducibility