$$5 \leq K \leq 10$$ is commonly used in Macine Learning. A too-small $K$ may lead to high bias, and a too-large $K$ may result in high variance. 

University of Michigan - Ann Arbor

To perform $$K$$-fold cross-validation on a dataset with $$n$$ observations, the training set is separated into $$K$$ sets by randomly sampling $$n/K$$ observations from the dataset into each fold. The model of interest is then trained on $$K-1$$ folds and validated on the left-out fold; that is, once the model is trained, we use it to make predictions on the group that was left out of training. This process repeats $$K$$ times. To evaluate the model's overall performance, we calculate the average validation error across all $$K$$ tests. $$K$$-fold cross-validation provides a robust estimate of the empirical testing error of a model and is particularly useful for selecting model designs and adjusting hyperparameters.

K-Fold Cross-Validation

James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical learning (Vol. 112, pp. 3-7). New York: springer.

An Introduction to Statistical Learning with Applications in R

K-Fold CV is usually the method of choice for CV, and functions as a sort of generalization of LOOCV, because:
1. It is much less computationally expensive than LOOCV.
2. It minimizes the **Bias-Variance Tradeoff**.

K-Fold Cross-Validation vs. Leave-One-Out Cross-Validation

$$5 \leq K \leq 10$$ in K-Fold Cross-Validation

When evaluating numerous hyperparameter configurations, the risk of multiple testing increases, leading to situations where validation performance appears favorable simply by chance rather than reflecting true generalization capability. However, when applied to a sufficiently large dataset with a standard range of hyperparameters, $$K$$-fold cross-validation tends to be reasonably resilient against this issue, providing a more robust estimate of true error than a single validation split.

Multiple Testing Resilience in K-Fold Cross-Validation

Throughout the training and model selection process, continuously monitoring both training and validation errors is crucial for diagnosing learning behavior. If a model exhibits an extremely low training error for a specific hyperparameter configuration while simultaneously yielding a considerably higher error during $$K$$-fold cross-validation, it serves as a strong indicator of overfitting. This divergence demonstrates that the model is memorizing the training data rather than generalizing well to unseen validation folds.

Identifying Overfitting via Cross-Validation

When selecting $$K$$ for $$K$$-Fold Cross-Validation, we must consider the bias-variance tradeoff. $$K$$-Fold Cross-Validation provides $$K$$ estimates of the model error. The mean of these errors indicates the model's bias, where a lower mean value implies higher accuracy. The model variance is determined by the standard deviation of these errors, where a lower variance indicates less fluctuation in model performance. Usually, it is difficult to reduce both model bias and variance simultaneously.

K-Fold Cross-Validation Bias-Variance Tradeoff

```python
from sklearn.model_selection import cross_val_score

cv_scores = cross_val_score(clf, X, y, cv=5)
```
Where:
- `clf`: The estimator/model object to evaluate.
- `X`: The input data (features) to fit.
- `y`: The target variable/labels.
- `cv`: The number of folds for cross-validation (defaults to 5).

The function returns an array of scores (such as accuracy) for each of the $$K$$ folds. The average performance of the model can be calculated using `cv_scores.mean()`.

sklearn.model_selection.cross_val_score

Cross-validation methods differ in the bias and variance of their estimated test error. A validation-set approach trains on a relatively small subset of the data and may therefore overestimate test error. Leave-one-out cross-validation trains each model on $$n-1$$ observations, producing an approximately unbiased estimate, but its estimates tend to have higher variance. In $$K$$-fold cross-validation, each model is trained on $$(K-1)n/K$$ observations, so values such as $$K=5$$ or $$K=10$$ generally provide an intermediate bias with lower variance than leave-one-out cross-validation.

Learn Before

Related