It depends on your algorithm and problem complexity. But here are practical guidelines.

Minimum Data by Algorithm

Linear/Logistic Regression: 10x features

• 10 features → 100 rows minimum
• 100 features → 1,000 rows minimum

Random Forest/Gradient Boosting: 10-50x features

• More forgiving with small data
• 1,000-10,000 rows is comfortable

Deep Learning: Thousands to millions

• Simple problems: 10,000+ rows
• Images: 100,000+ images per class
• NLP: Millions of examples

Quality Beats Quantity

1,000 clean, relevant examples > 100,000 noisy ones.

Good data:

• Representative of real-world cases
• Correctly labeled
• Balanced classes
• Relevant features

What If You Don't Have Enough Data?

Transfer learning: Use pre-trained models
Data augmentation: Create variations (rotate images, synonym replacement)
Simpler models: Use algorithms that need less data
Collect more: Sometimes you just need more data

The Rule of Thumb

For most business problems:
• 1,000 rows: Can try ML
• 10,000 rows: Comfortable for tree-based models
• 100,000+ rows: Neural networks become viable