Unraveling overoptimism and publication bias in ML-driven science

Machine Learning (ML) is increasingly used across many disciplines with impressive reported results across many domain areas. However, recent studies suggest that the published performance of ML models are often overoptimistic. Validity concerns are underscored by findings of an inverse relationship between sample size and reported accuracy in published ML models, contrasting with the theory of learning curves where accuracy should improve or remain stable with increasing sample size. This paper investigates factors contributing to overoptimistic accuracy reports in ML-driven science, focusing on data leakage and publication bias. We introduce a novel stochastic model for observed accuracy, integrating parametric learning curves and the aforementioned biases. We then construct an estimator that corrects for these biases in observed data. Theoretical and empirical results show that our framework can estimate the underlying learning curve, providing realistic performance assessments from published results. Applying the model to meta-analyses in ML-driven science, including neuroimaging-based and speech-based classifications of neurological conditions, we find prevalent overoptimism and estimate the inherent limits of ML-based prediction in each domain.