SHAP tells you one prediction; FACET tells you the whole model

What it does

FACET is a Python library for explaining supervised ML models at the global level, not just per-prediction. It ships two main ideas: model inspection, which quantifies pairwise dependencies and interactions between features, and model simulation, which runs “virtual experiments” on a fitted model to find scenarios that optimise predicted outcomes. It also wraps scikit-learn’s pipeline workflow with model-selection tooling and its companion library sklearndf, which keeps pandas DataFrames traceable end to end.

The interesting bit

The inspection piece introduces two asymmetric, percentage-based metrics for every feature pair: synergy (how much a feature leans on another to make a prediction) and redundancy (how much one feature duplicates another’s information). The asymmetry is the clever part — house size “knows” more of what bedroom count does than vice versa, and the metrics capture that direction. This complements SHAP’s local explanations with a global view, and there’s a formal write-up on arXiv (2107.12436).

Key highlights

• Synergy and redundancy matrices for all feature pairs, visualised out of the box
• Model simulation with bootstrapping (stationary and stratified) to quantify uncertainty in virtual experiments
• LearnerSelector ranks competing hyperparameter configurations and even different learners side by side
• Built on scikit-learn pipelines via sklearndf, so feature names survive the whole workflow
• Available on both PyPI and conda

Caveats

• Scikit-learn estimators only — this is not for deep learning frameworks
• The primary language on GitHub is Jupyter Notebook, so expect the repo to lean heavily on tutorial notebooks rather than a compact library codebase

Verdict

Worth a look if you fit scikit-learn models and need to explain feature relationships to humans, not just individual predictions. If you’re outside the sklearn/pandas world, it’s not for you.