fippy: Feature Importance in Python 🐬

A Python package for model-agnostic feature importance with statistical inference. Fippy implements a unified framework where feature importance methods are composed from three orthogonal axes:

Axis	Options	Description
Attribution	loo, shapley	Leave-one-out vs. Shapley value attribution
Restriction	resample, marginalize, refit	How the removed feature is handled
Distribution	marginal, conditional	Which distribution replacement values are drawn from

This gives rise to well-known methods as special cases:

Method	Attribution	Restriction	Distribution	Alias
PFI	loo	resample	marginal	pfi()
CFI	loo	resample	conditional	cfi()
RFI	loo	resample	conditional + G	rfi()
LOCO	loo	refit	—	loco()
SAGE	shapley	marginalize	marginal or conditional	sage()

Installation

Requires Python >= 3.9.

pip install -e path/to/fippy

Quick start

import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split

from fippy import Explainer
from fippy.samplers import GaussianSampler
from fippy.losses import squared_error

# Prepare data and model
X, y = ...  # your data as pd.DataFrame and array
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

model = RandomForestRegressor()
model.fit(X_train, y_train)

# Create explainer
explainer = Explainer(model.predict, X_train, loss=squared_error)

# Permutation Feature Importance (marginal)
result_pfi = explainer.pfi(X_test, y_test, n_repeats=10)
result_pfi.importance()

Conditional methods

For conditional feature importance, pass a sampler:

sampler = GaussianSampler(X_train)
explainer = Explainer(model.predict, X_train, loss=squared_error, sampler=sampler)

# Conditional Feature Importance
result_cfi = explainer.cfi(X_test, y_test, n_repeats=10)
result_cfi.importance()

Refit-based methods

For LOCO, pass a learner (unfitted sklearn estimator) and set training labels:

explainer.set_y_train(y_train)
result_loco = explainer.loco(X_test, y_test, learner=RandomForestRegressor())
result_loco.importance()

SAGE (Shapley-based)

result_sage = explainer.sage(X_test, y_test, distribution="marginal", n_samples=50)
result_sage.importance()

Key parameters

restriction — How the dropped feature is handled:

• "resample": Replace the feature with a single random draw from a reference distribution. Fast, but introduces sampling noise.
• "marginalize": Replace the feature with n_samples random draws and average the predictions. Approximates the expected prediction over the reference distribution. More accurate than resample but slower.
• "refit": Retrain the model without the feature entirely. No sampling involved; measures importance through model performance change.

distribution — Which reference distribution replacement values are drawn from (for resample and marginalize):

• "marginal": Draw from the unconditional distribution P(X_j). Breaks dependence between the dropped feature and all others. Used by PFI.
• "conditional": Draw from P(X_j | X_{-j}), respecting the dependence structure. Requires a sampler. Used by CFI/RFI.

Not applicable for restriction="refit".

n_repeats — Number of times the importance computation is repeated with fresh perturbed samples for the dropped features. The variance across repeats captures Monte Carlo noise and is used by ci() for confidence intervals. Higher values give more stable estimates.

n_samples — Number of replacement samples drawn per observation when using restriction="marginalize". The predictions are averaged over these samples to approximate the expectation. Required for sage() and loo(..., restriction="marginalize"). Not used for resample (which draws a single sample) or refit.

n_permutations — Maximum number of random feature orderings for Shapley value estimation (default: 500). Shapley values are approximated by averaging marginal contributions over random permutations. Computation stops early if the estimates converge (controlled by convergence_threshold).

n_jobs — Number of parallel threads for the feature loop in LOO methods (default: 1). Uses joblib with thread-based parallelism. Set to -1 to use all cores. Not supported for Shapley.

Using the generic interface

All convenience methods are shortcuts for loo() and shapley():

# Equivalent to explainer.pfi(X_test, y_test)
result = explainer.loo(X_test, y_test, "resample", distribution="marginal")

Statistical inference

ExplanationResult provides built-in inference on the importance scores.

# Mean importance with standard deviation
result.importance()

# Confidence intervals (t-based or quantile-based)
result.ci(alpha=0.05, method="t")

# Hypothesis test H0: importance_j <= 0
result.test(method="t", alternative="greater", p_adjust="bonferroni")

# Observation-wise importance scores
result.obs_importance()

# Relative importance (as fraction of baseline loss)
result.importance(relative=True)

Multiple testing corrections: "bonferroni", "holm", "bh" (Benjamini-Hochberg).

Plotting

Visualize results with horizontal bar plots showing confidence intervals:

# Single plot
result.hbarplot(figsize=(8, 4))

# Side-by-side comparison
import matplotlib.pyplot as plt
fig, axes = plt.subplots(1, 2, figsize=(14, 4))
result_pfi.hbarplot(ax=axes[0])
result_cfi.hbarplot(ax=axes[1])
fig.tight_layout()

See Example.ipynb for a complete walkthrough with plots.

Feature groups

Features can be grouped and assessed jointly:

# As a dict
explainer.pfi(X_test, y_test, features={"size": ["width", "height"], "color": ["r", "g", "b"]})

# As a list (each element becomes one group)
explainer.pfi(X_test, y_test, features=["width", "height", "color"])

Samplers

Sampler	Distribution	Description
PermutationSampler	marginal	Draws from training marginal (used automatically)
GaussianSampler	conditional	Multivariate Gaussian conditional P(X_J | X_S)

Additional samplers (regression-based, ARF, TabPFN) are planned.

Serialization

result.to_csv("importance.csv")
from fippy import ExplanationResult
loaded = ExplanationResult.from_csv("importance.csv")

Status

The package is under active development. The core API is stable, but additional samplers and cross-validation features are planned.

References

The package was previously called rfi and accompanies our paper on Relative Feature Importance: [arXiv]