Leakage and the Reproducibility Crisis in ML-based Science

We argue that there is a reproducibility crisis in ML-based science. We compile evidence of this crisis across fields, identify data leakage as a pervasive cause of reproducibility failures, conduct our own reproducibility investigations using in-depth code-review, and propose a solution.

Context

Many quantitative science fields are adopting the paradigm of predictive modeling using machine learning. We welcome this development. At the same time, as researchers whose interests include the strengths and limits of machine learning, we have concerns about reproducibility and overoptimism.

There are many reasons for caution:

  • Performance evaluation is notoriously tricky in machine learning.
  • ML code tends to be complex and as yet lacks standardization.
  • Subtle pitfalls arise from the differences between explanatory and predictive modeling.
  • The hype and overoptimism about commercial AI may spill over into ML-based scientific research.
  • Pressures and publication biases that have led to past reproducibility crises are also present in ML-based science.
Given these reasons, we view reproducibility difficulties as the expected state of affairs until best practices become better established and understood. The spate of reproducibility failures (that we have compiled below) highlights the immaturity of ML-based science, the critical need for ongoing work on methods and best practices, and the importance of treating the results from this body of work with caution.



Scope

We focus on reproducibility issues in ML-based science, which involves making a scientific claim using the performance of the ML model as evidence. There is a much better known reproducibility crisis in research that uses traditional statistical methods. We also situate our work in contrast to other ML domains, such as methods research (creating and improving widely-applicable ML methods), ethics research (studying the ethical implications of ML methods), engineering applications (building or improving a product or service), and modeling contests (improving predictive performance on a fixed dataset created by an independent third party). Investigating the validity of claims in all of these areas is important, and there is ongoing work to address reproducibility issues in these domains.

Various domains in which ML methods are used. In our work, we focus on ML-based science.

A non-exhaustive categorization of focus areas in ML literature. In our work, we focus on ML-based science.

The running list below consists of papers that highlight reproducibility failures or pitfalls in ML-based science. We find 20 papers from 17 fields where errors have been found, collectively affecting 329 papers and in some cases leading to wildly overoptimistic conclusions. In each case, data leakage causes errors in the modeling process.

Field Paper Year Num. papers reviewed Num. papers w/pitfalls Pitfalls
Medicine Bouwmeester et al. 2012 71 27 No train-test split
Neuroimaging Whelan et al. 2014 14 No train-test split; Feature selection on train and test set
Autism Diagnostics Bone et al. 2015 3 Duplicates across train-test split; Sampling bias
Bioinformatics Blagus et al. 2015 6 Pre-processing on train and test sets together
Nutrition research Ivanescu et al. 2016 4 No train-test split
Software engineering Tu et al. 2018 58 11 Temporal leakage
Toxicology Alves et al. 2019 1 Duplicates across train-test split
Satelitte imaging Nalepa et al. 2019 17 17 Non-independence between train and test sets
Clinical epidemiology Christodoulou et al. 2019 71 48 Feature selection on train and test set
Tractography Poulin et al. 2019 4 2 No train-test split
Brain-computer interfaces Nakanishi et al. 2020 1 No train-test split
Histopathology Oner et al. 2020 1 Non independence between train and test sets
Computer security Arp et al. 2020 30 30 No train-test split; Pre-processing on train and test sets together; Illegitimate features; others
Neuropsychiatry Poldrack et al. 2020 100 53 No train-test split; pre-processing on train and test sets together
Medicine Vandewiele et al. 2021 24 21 Feature selection on train-test sets; Non-independence between train and test sets; Sampling bias
Radiology Roberts et al. 2021 62 62 No train-test split; duplicates in train and test sets; sampling bias
IT Operations Lyu et al. 2021 9 3 Temporal leakage
Medicine Filho et al. 2021 1 Illegitimate features
Neuropsychiatry Shim et al. 2021 1 Feature selection on training and test sets
Genomics Barnett et al. 2022 41 23 Feature selection on training and test sets

Data leakage has long been recognized as a leading cause of errors in ML applications. In formative work on leakage, Kaufman et al. provide an overview of different types of errors and give several recommendations for mitigating these errors. Since this paper was published, the ML community has investigated the impact of leakage in several engineering applications and modeling competitions. However, leakage occurring in ML-based science has not been comprehensively investigated. As a result, mitigations for data leakage in scientific applications of ML remain understudied.

A taxonomy of data leakage can enable a better understanding of why leakage occurs in ML-based science and inform potential solutions. We present a fine-grained taxonomy of 8 types of leakage that range from textbook errors to open research problems. Our taxonomy is comprehensive and addresses data leakage arising during the data collection, pre-processing, modeling and evaluation steps. In particular, our taxonomy addresses all cases of data leakage that we found in our survey. We provide an overview of the types of leakage here, a more detailed taxonomy is included in our paper.

1. Lack of clean separation of training and test set: If the training dataset is not separated from the test dataset during all pre-processing, modeling and evaluation steps, the model has access to information in the test set before its performance is evaluated.

2. Model uses features which are not legitimate: The model has access to features that should not be legitimately available for use in the modeling exercise, for instance if they are a proxy for the outcome variable.

3. Test set is not drawn from the distribution of interest: The distribution of data on which the performance of an ML model is evaluated differs from the distribution of data about which the scientific claims are made.

Our taxonomy of data leakage highlights several failure modes which are prevalent in ML-based science. To address leakage, researchers using ML methods need to connect the performance of their ML models to their scientific claims. To detect cases of leakage, we provide a template for a model info sheet which should be included when making a scientific claim using predictive modeling. The template consists of precise arguments needed to justify the absence of leakage, and is inspired by Mitchell et al.'s model cards for increasing the transparency of ML models.

Model info sheets can be voluntarily used by researchers to detect leakage, or journals could encourage or require authors to provide them. Of course, model info sheets can’t prevent researchers from making false claims, but we hope they can make errors more apparent. Note that for model info sheets to be verified, the analysis must be computationally reproducible. Also, model info sheets don’t address reproducibility issues other than leakage. The model info sheet template is currently a beta version. We welcome feedback and are continuing to make changes to the template based on feedback.

We find that prominent studies on civil war prediction claiming superior performance of ML models over baseline Logistic Regression models fail to reproduce. Our results provide two reasons to be skeptical of the use of ML methods in this research area, by both questioning their usefulness and highlighting the pitfalls of applying them correctly. While none of these errors could have been caught by reading the papers, our model info sheets enable the detection of leakage in each case.

 A comparison of reported results vs. corrected results in the 4 papers on civil war prediction that compare the performance of ML models and Logistic Regression models.

A comparison of reported and corrected results in civil war prediction papers published in top Political Science journals. The main findings of each of these papers are invalid due to various forms of data leakage: Muchlinski et al. impute the training and test data together, Colaresi & Mahmood and Wang incorrectly reuse an imputed dataset, and Kaufman et al. use proxies for the target variable which causes data leakage. The use of model info sheets would detect leakage in every paper. When we correct these errors, complex ML models (such as Adaboost and Random Forests) do not perform substantively better than decades-old Logistic Regression models for civil war prediction in each case. Each column in the table outlines the impact of leakage on the results of a paper.

Reproduction materials on CodeOcean          List of papers in our systematic review

We acknowledge that there isn't consensus about the term reproducibility, and there have been a number of recent attempts to define the term and create consensus. One possible definition is computational reproducibility — when the results in a paper can be replicated using the exact code and dataset provided by the authors. We argue that this definition is too narrow because even cases of outright bugs in the code would not be considered irreproducible under this definition. Therefore we advocate for a standard where bugs and other errors in data analysis that change or challenge a paper's findings constitute irreproducibility. We elaborate this perspective here.

Reproducibility failures don’t mean a claim is wrong, just that evidence presented falls short of the accepted standard or that the claim only holds in a narrower set of circumstances than asserted. We don’t view reproducibility failures as signs that individual authors or teams are careless, and we don’t think any researcher is immune. One of us (Narayanan) has had multiple such failures in his applied-ML work and expects that it will probably happen again.

We call it a crisis for two related reasons. First, reproducibility failures in ML-based science are systemic. In nearly every scientific field that has carried out a systematic study of reproducibility issues, papers are plagued by common pitfalls. In many systematic reviews, a majority of the papers reviewed suffer from these pitfalls. Second, despite the urgency of addressing reproducibility failures, there aren’t yet any systemic solutions.

To cite this work, please use this BibTeX entry. This is a project by Sayash Kapoor and Arvind Narayanan. We are researchers in the department of computer science and the Center for Information Technology Policy at Princeton University.

Our interest in this topic arose during a graduate seminar on Limits to Prediction. Narayanan offered this course together with Prof. Matthew Salganik in Fall 2020, and Kapoor took the course. The course aimed to critically examine the narrative about the ability to predict the future with ever-increasing accuracy given bigger datasets and more powerful algorithms. The work on reproducibility pitfalls is one aspect of our broader interest in limits to prediction.