Skip to content

SWIFT

SHAP-Weighted Impact Feature Testing for Model-Aware Distribution Monitoring.

swift answers: which feature distributions have shifted in a way that actually matters to my model? Traditional drift detection (KS test, PSI, …) flags any statistical shift, regardless of whether the model cares. SWIFT compares SHAP-transformed distributions between reference and monitoring data, weighting every distribution change by its impact on model predictions — so only the shifts that affect model behavior are flagged.

from swift import SWIFTMonitor

# Create monitor with a trained tree-ensemble model
monitor = SWIFTMonitor(model=lgb_model, n_permutations=200)

# Fit on reference data (stages 1-3)
monitor.fit(X_ref)

# Test monitoring data for drift (stages 4-5)
result = monitor.test(X_mon)
print(result.drifted_features)
SWIFT Scores — drifted data
SWIFT scores per feature — drifted features highlighted in red

Highlights

  • Model-aware by construction — feature values are mapped to the mean SHAP value of the model-derived bucket they fall in, so regions the model ignores cannot raise an alarm (how it works).
  • Buckets come from the model itself — bucket boundaries are the split thresholds of the trained LightGBM or XGBoost ensemble, not arbitrary quantiles (decision points, buckets).
  • Statistical rigor — Wasserstein distances on SHAP-transformed distributions, permutation-test p-values, and Benjamini-Hochberg or Bonferroni multiple testing correction (testing).
  • scikit-learn compatible — SWIFTMonitor inherits from BaseEstimator and TransformerMixin: fit / transform / score / test, get_params / set_params, and clone all work (scikit-learn integration).
  • See the drift — per-feature bucket profiles (SHAP response curve + observation density) and drift-colored score plots, all returning (Figure, Axes) for customization (visualization).
  • Sample vs. sample mode — compare any two data windows against each other with X_compare, reusing the SHAP transformation learned from the reference (FAQ).

Where to start

  1. Getting started — install and your first drift test in five minutes.
  2. How it works — the full five-stage pipeline, from split thresholds to corrected p-values.
  3. Quickstart tutorial — a runnable notebook, from training a model to reading the drift verdict.
  4. API reference — every public class and function.

Citation

@article{swift2025,
  title={SWIFT: SHAP-Weighted Impact Feature Testing for Model-Aware Distribution Monitoring},
  year={2025}
}