Optimizing Sample Size in Machine Learning: Balancing Complexity and Generalization

This paper introduces a new direction in estimating an optimal sample size for a binary classification problem using two theoretical frameworks, namely Rademacher complexity and the PAC-Bayesian generalization bound.First, the Rademacher complexity provides a measure of the capacity of a hypothesis class, characterizing its risk of overfitting, while the second-the PAC-Bayesian generalization bound-provides a probabilistic upper bound on the generalization error from a Bayesian viewpoint.We unify these two bounds to get a tighter generalization error bound that will enable a closer estimate of the optimal sample size in achieving model performance that one can count on.The experimental results showed that the combination of both bounds allows a more thorough understanding of the generalization performance of the classifiers.