aerdem4/lofo-importance

Iteratively removes each feature and retrains the model across validation folds to measure performance impact, providing model-agnostic importance scores that account for feature interactions and generalize to unseen data. Supports custom validation schemes, custom models (defaults to LightGBM), and feature grouping for high-dimensional features; also includes FLOFO, a faster permutation-based variant that groups samples during shuffling to avoid unrealistic feature value replacements.

863 stars and 1,583 monthly downloads. No commits in the last 6 months. Available on PyPI.