From theory to practice: implementing and evaluating e-fold cross-validation

In this paper, we present e-fold cross-validation, an energy-efficient alternative to k-fold, which dynamically adjusts the number of folds to reduce computational resources while ensuring accurate performance estimates. This dynamic adjustment is implemented by a stopping criterion, which checks after each fold whether the standard deviation of the previous evaluated folds has repeatedly decreased or has not changed significantly. Once the stopping criterion is met, the e-fold cross-validation is stopped early. We evaluated e-fold cross-validation on 15 datasets and 10 machine-learning algorithms. Compared to 10-fold cross-validation, e-fold cross-validation needed about 4 folds less for comparable results, resulting in savings of about 40% evaluation time, computational resources and energy consumption. The percentage difference in performance metrics between e-fold and standard 10-fold cross-validation with larger datasets was generally less than 2%. More powerful models show even smaller discrepancies. Statistical significance was confirmed with 96% of iterations falling within the confidence interval for all algorithms tested. Our results indicate that e-fold cross-validation is a robust and efficient alternative to the conventional k-fold cross-validation providing reliable performance evaluation at lower computational costs.