Leakly 0.1.2

Leakage checks for machine-learning pipelines using permutation tests.

Leakly: Leakage checks for any machine-learning pipeline

Leakly uses label permutation to test whether a machine-learning pipeline performs above chance when no true signal is present.

Above-chance performance after permutation may indicate leakage from preprocessing, feature selection, tuning, or another step of the pipeline.

How it works

1. Permute labels to remove the real feature-label association.
2. Run the full pipeline exactly as in the original analysis.
3. Compare the permuted score distribution with chance level.
4. Above-chance permuted performance suggests possible leakage.

Leakly includes example configurations for a leaky pipeline and a non-leaky pipeline so users can inspect the effect directly.

Install

pip install Leakly

Quick Start on Colab:

Key Python snippet

from leakly import (
    MLPipeline,
    SummaryPlotter,
    load_example_leakage_config,
    permute_label)

scores = []
for seed in range(100):
    permuted_y = permute_label(y, random_state=seed)
    score = (
        # Replace with any user-defined pipeline
        MLPipeline(
            X,
            permuted_y,
            covariates=covariates,
            config=load_example_leakage_config(),
        ).fit()
    ).evaluate()
    scores.append(score)

SummaryPlotter(scores, chance_level=0.5).plot()

FAQ

Can Leakly check my own pipeline?

Yes. Leakly can evaluate any pipeline that takes X, y, optional covariates, and returns a test score. The key is to run the full pipeline exactly as in the real analysis, including preprocessing, feature selection, tuning, and evaluation.

Why can a leaky pipeline score well on permuted labels?

If leakage occurs, information from test samples can enter the analysis before the train/test split or outside the cross-validation loop. Common sources include feature selection, scaling, imputation, covariate adjustment, dimensionality reduction, or hyperparameter tuning performed on all samples.

In high-dimensional data such as omics and neuroimaging, random features can appear predictive by chance. If a pipeline can retain these spurious patterns, it may perform above chance even after labels are permuted.

How many permutations should I run?

Use 100 for a quick check. Use 1,000 or more for publication-level evidence.

License

MIT. See LICENSE.