# Leakage and the Reproducibility Crisis in ML-based Science

**Authors:** Sayash Kapoor, Arvind Narayanan  
**arXiv:** [2207.07048](https://arxiv.org/abs/2207.07048)  
**Date:** 2022-07-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.07048v1/

## Key points

- Machine learning (ML) methods are widely used in scientific fields for prediction and forecasting.
- Data leakage is a significant issue that can lead to reproducibility failures in ML-based science.
- The authors conducted a systematic investigation and found data leakage to be prevalent in 17 fields, affecting 329 papers.
- They propose a fine-grained taxonomy of eight types of leakage to understand the problem better.
- The authors argue for fundamental methodological changes to prevent cases of leakage from being published.
- Model info sheets are proposed as a reporting mechanism to address all types of leakage and detect instances before publication.
- A reproducibility study in civil war prediction shows that papers claiming superior performance by complex ML models fail to reproduce due to data leakage.
- Complex ML models do not significantly outperform older statistical models like Logistic Regression (LR).
- Model info sheets could have helped detect these errors, emphasizing their importance as a reporting standard.

## Summary

In the paper "Leakage and the Reproducibility Crisis in ML-based Science," authors Sayash Kapoor and Arvind Narayanan address the widespread use of machine learning (ML) methods for prediction and forecasting in various scientific fields. They highlight the methodological pitfalls, particularly data leakage, that can lead to severe reproducibility failures in ML-based science. The authors conduct a systematic investigation into reproducibility issues in ML-based science and find that data leakage is indeed a prevalent problem. They analyze literature from research communities that have adopted ML methods and identify 17 fields where errors related to data leakage have been found. These errors collectively affect 329 papers and, in some cases, result in overly optimistic conclusions. To provide a comprehensive understanding of data leakage, the authors present a fine-grained taxonomy of eight types of leakage ranging from textbook errors to open research problems. They argue for fundamental methodological changes to ML-based science to prevent cases of leakage from being published. To address this issue, Kapoor and Narayanan propose model info sheets as a reporting mechanism for scientific claims based on ML models. These info sheets aim to address all types of leakage identified in their survey by providing detailed information about the models used and potential sources of data leakage which can help detect instances of leakage before publication. To investigate the impact of reproducibility errors and evaluate the effectiveness of model info sheets, the authors undertake a reproducibility study in the field of civil war prediction. In this field, complex ML models are believed to outperform older statistical models such as Logistic Regression (LR). However, their study reveals that all papers claiming superior performance by complex ML models compared to LR models fail to reproduce due to data leakage. Furthermore, they find that complex ML models do not significantly outperform decades-old LR models. The authors emphasize that these errors could not have been caught by simply reading the papers but could have been detected through the use of model info sheets. They conclude that adopting model info sheets as a reporting standard can help identify and prevent data leakage, ultimately improving the reproducibility of ML-based science.