cross-validation

A single train-test split gives a noisy estimate of model performance — different splits give different results, especially on small datasets.

A resampling technique that repeatedly splits data into train and validation sets to get a more robust estimate of model performance.

1. Choose k (typically 5 or 10)
2. Split data into k roughly equal folds
3. For each fold i: train on all folds except i, validate on fold i
4. Compute performance metric for each fold
5. Average the k performance scores for final estimate

• Robust: reduces variance of performance estimate compared to single split
• Computationally expensive: requires training k models instead of one
• Data-efficient: uses all data for both training and validation (just not simultaneously)
• Standard: the gold standard for model evaluation in ML research

• Built from: Train-Test Split — cross-validation is a multi-split extension
• Related: Overfitting — CV helps detect overfitting more reliably
• Related: Supervised Learning — CV is used to evaluate supervised models
• Related: Data Modeling — model selection uses CV to compare algorithms

• Time series data: standard k-fold breaks temporal ordering (use time-series CV)
• Data leakage: preprocessing (normalization) must be done inside each fold
• Stratification: for classification, ensure each fold has representative class proportions
• Small datasets: leave-one-out CV (k=n) can have high variance despite being “exact”