Local Feature Attributions

Local attributions explain individual predictions by decomposing them into feature contributions.

scikit-explain supports:

• SHAP — SHapley Additive Explanations

• LIME — Local Interpretable Model Explanations

• Tree Interpreter — tree-based model decomposition

import skexplain
import plotting_config
import shap

# Loading the training data and pre-fit models
estimators = skexplain.load_models()
X, y = skexplain.load_data()

# Pick a single example for local attribution
single_example = X.iloc[[0]]

explainer = skexplain.ExplainToolkit(estimators[0], X=single_example)
explainer.set_plotting_config(
    display_feature_names=plotting_config.display_feature_names,
    display_units=plotting_config.display_units,
)

Computing Local Attributions

# SHAP masker handles missing features using correlations in the dataset
shap_kws = {
    'masker': shap.maskers.Partition(X, max_samples=100, clustering='correlation'),
    'algorithm': 'permutation',
}

# LIME requires the training data
lime_kws = {
    'training_data': X.values,
    'categorical_names': ['rural', 'urban'],
}

# Compute all three attribution methods at once
contrib_ds = explainer.local_attributions(
    method=['shap', 'lime', 'tree_interpreter'],
    shap_kws=shap_kws,
    lime_kws=lime_kws,
)

# The result is an xarray.Dataset with attribution values for each method
contrib_ds

Plotting Attributions (Waterfall Plot)

fig, axes = explainer.plot_contributions(contrib=contrib_ds)

Performance-Based Attributions

Average attributions across the best and worst-performing examples.

# Create a new ExplainToolkit with the full dataset
explainer = skexplain.ExplainToolkit(estimators[0], X=X, y=y)
explainer.set_plotting_config(
    display_feature_names=plotting_config.display_feature_names,
    display_units=plotting_config.display_units,
)

# Compute performance-based attributions using tree interpreter
tree_results = explainer.average_attributions(
    method='tree_interpreter',
    performance_based=True,
    n_samples=100,
)

# Plot the top 5 features for selected performance categories
perf_keys = ['Best Hits', 'Worst False Alarms', 'Worst Misses']

fig, axes = explainer.plot_contributions(
    contrib=tree_results,
    perf_keys=perf_keys,
    num_features=5,
)

Regression Example

from sklearn.datasets import fetch_california_housing
from sklearn.ensemble import RandomForestRegressor

data = fetch_california_housing()
X_reg = data['data']
y_reg = data['target']
feature_names = data['feature_names']

model = RandomForestRegressor()
model.fit(X_reg, y_reg)

single_example = X_reg[[0]]

explainer = skexplain.ExplainToolkit(
    ('Random Forest', model),
    X=single_example,
    feature_names=feature_names,
)

shap_kws = {
    'masker': shap.maskers.Partition(X_reg, max_samples=100, clustering='correlation'),
    'algorithm': 'auto',
}

results = explainer.local_attributions(method='shap', shap_kws=shap_kws)
fig = explainer.plot_contributions(results, figsize=(4, 8))