Definition
- Multiclass Stratified CV is a cross-validation method for multiclass classification problems where each fold maintains approximately the same class distribution as the original dataset.
- It’s an extension of Stratified CV for binary classification to cases with more than two classes.
How It Works
- Calculate the overall class distribution in the dataset.
- Example: Class A = 60%, Class B = 30%, Class C = 10%.
- When splitting into k folds, ensure each fold has roughly the same ratio.
- Each fold should contain A: 60%, B: 30%, C: 10%.
- Guarantees that both training and validation sets are representative of all classes.
Example
- Dataset = 1,000 samples (A = 600, B = 300, C = 100).
- k = 5 folds.
- Regular k-fold: some folds may have very few Class C samples.
- Multiclass Stratified k-fold:
- Each fold → A = 120, B = 60, C = 20 → balanced like the original distribution.
Why It Matters
- In imbalanced multiclass datasets, regular k-fold may underrepresent minority classes in some folds.
- Multiclass stratification ensures fair and stable evaluation of the model.
- Prevents misleading performance metrics (e.g., if one fold has almost no samples of a class).
When to Use
- Any multiclass classification problem (3+ classes).
- Especially important when class imbalance exists.
- Not used for regression problems (stratification applies only to classification).
Summary
Multiclass Stratified CV = k-fold cross-validation where each fold preserves the class distribution across multiple classes.
- Ensures every fold contains all classes in the right proportions.
- Provides more reliable and fair model evaluation.