Leakage and the reproducibility crisis in machine-learning-based science

Area of Science:

• Quantitative sciences
• Computational social science

Background:

• Machine-learning (ML) methods are increasingly used in quantitative sciences.
• Methodological pitfalls, such as data leakage, can compromise the reliability of ML-based research.
• Reproducibility issues are a significant concern in the application of ML in science.

Purpose of the Study:

• To systematically investigate reproducibility issues in ML-based science, focusing on data leakage.
• To identify and categorize different types of data leakage in ML research.
• To propose a method for testing and mitigating data leakage.

Main Methods:

• Systematic literature survey across 17 scientific fields that utilize ML methods.
• Development of a detailed taxonomy of eight types of data leakage.
• Introduction of model info sheets for researchers to test for leakage.
• Reproducibility study comparing complex ML models with traditional logistic regression (LR) in civil war prediction.

Main Results:

• Data leakage was identified in 294 papers across 17 fields, often resulting in exaggerated conclusions.
• A comprehensive taxonomy of eight data leakage types was established.
• Model info sheets were proposed as a tool for assessing leakage.
• In a civil war prediction case study, corrected ML models showed no substantial performance advantage over LR models.

Conclusions:

• Data leakage is a widespread problem in ML-based science, undermining reproducibility and leading to inflated performance metrics.
• The proposed taxonomy and model info sheets offer a framework for identifying and addressing leakage.
• Complex ML models may not inherently outperform simpler, established statistical methods like LR when methodological rigor is applied.