SWIFT: SHAP-Weighted Impact Feature Testing

Model-aware distribution monitoring for production ML systems.

SWIFT detects feature drift by comparing SHAP-transformed distributions between reference and monitoring data. Unlike traditional drift detection that treats features independently of the model, SWIFT weights distribution changes by their impact on model predictions — flagging only the shifts that affect model behavior.

Installation

pip install swift-monitoring

Or with uv:

uv pip install swift-monitoring

For development (from source):

git clone https://github.com/wlazlod/swift.git
cd swift
uv pip install -e ".[dev]"

Requires Python >= 3.11.

Quick Start

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)

The SWIFT Pipeline

SWIFT executes a 5-stage pipeline:

Stage	Operation	Description
1	Extraction	Extract decision points (split thresholds) from the trained model
2	Bucketing	Partition each feature's domain into buckets based on decision points
3	SHAP Normalization	Compute bucket-level mean SHAP values on reference data
4	Wasserstein Distance	Measure distance between SHAP-transformed distributions
5	Permutation Test + MTC	Estimate p-values and apply multiple testing correction

Stages 1-3 run during fit(). Stages 4-5 run during test().

API Reference

SWIFTMonitor

SWIFTMonitor(
    model,                          # Trained tree-ensemble (LightGBM/XGBoost)
    order=1,                        # Wasserstein order (1 or 2)
    n_permutations=1000,            # Permutations for p-value estimation
    alpha=0.05,                     # Significance level
    correction="benjamini-hochberg",# MTC method ("bonferroni", "bh", "fdr")
    n_synthetic=10,                 # Synthetic observations for empty buckets
    max_samples=None,               # Max pool size for permutation test
    random_state=42,                # RNG seed
)

Core Methods

Method	Input	Output	Description
fit(X)	DataFrame	self	Learn reference distribution (stages 1-3)
transform(X)	DataFrame	DataFrame	Map values to bucket-level mean SHAP
fit_transform(X)	DataFrame	DataFrame	Fit and transform in one call
score(X)	DataFrame	dict[str, float]	Per-feature Wasserstein distances vs. reference (stage 4)
score(X, X_compare=Y)	DataFrame, DataFrame	dict[str, float]	Per-feature distances: X vs. Y (SHAP transform from reference)
test(X)	DataFrame	SWIFTResult	Full drift test vs. reference with p-values and MTC
test(X, X_compare=Y)	DataFrame, DataFrame	SWIFTResult	Full drift test: X vs. Y (SHAP transform from reference)

Visualization Methods

Method	Description
plot_buckets(feature_name)	SHAP response curve + reference density per bucket
plot_buckets(feature_name, X=...)	Custom density source (e.g. monitoring sample instead of reference)
plot_buckets(feature_name, X_compare=...)	Side-by-side density comparison
plot_buckets(feature_name, x_axis="natural")	Feature-value scale on x-axis (instead of bucket indices)
plot_swift_scores(result)	Bar chart of SWIFT scores, colored by drift status
plot_swift_scores(result, result_compare=...)	Grouped bars comparing two test results

Both return (Figure, Axes) tuples for further customization.

Result Objects

SWIFTResult — returned by test():

Attribute	Type	Description
feature_results	tuple[FeatureSWIFTResult, ...]	Per-feature results
swift_max	float	Maximum SWIFT score across features
swift_mean	float	Mean SWIFT score across features
num_drifted	int	Number of drifted features
drifted_features	tuple[str, ...]	Names of drifted features

FeatureSWIFTResult — per-feature detail:

Attribute	Type	Description
feature_name	str	Feature name
swift_score	float	SWIFT score (Wasserstein distance)
p_value	float	Permutation test p-value
is_drifted	bool	Drift flag after MTC
num_buckets	int	Number of buckets used

Visualization

Bucketing Profile

# Single sample — reference density only
fig, ax = monitor.plot_buckets("feature_0")

# Comparison — reference vs. drifted density side-by-side
fig, ax = monitor.plot_buckets(
    "feature_0",
    X_compare=X_drifted,
    labels=("Reference", "Drifted"),
)

# Natural x-axis — bucket midpoints on the feature-value scale
fig, ax = monitor.plot_buckets("feature_0", x_axis="natural")

# Custom density source — show monitoring density instead of reference
fig, ax = monitor.plot_buckets("feature_0", X=X_mon)

The bucketing profile shows:

• Left y-axis: Mean SHAP per bucket (line + markers) with a shaded 95% band (mean +/- 2 std)
• Right y-axis: Observation density (line + fill, fraction of samples in each bucket)
• X-axis: Bucket intervals (default), compact labels B0, B1, ... for >20 buckets, or feature-value scale with x_axis="natural"
• SHAP=0 reference: Dashed grey horizontal line

SWIFT Scores

# Single result — bars colored by drift status
fig, ax = monitor.plot_swift_scores(result, threshold=0.01)

# Comparison — grouped bars from two test results
fig, ax = monitor.plot_swift_scores(
    result_clean,
    result_compare=result_drifted,
    labels=("Clean", "Drifted"),
)

The score plot shows:

• Single mode: Bars colored red (drifted) or blue (not drifted), with horizontal lines for SWIFT max, SWIFT mean, and an optional threshold
• Comparison mode: Grouped side-by-side bars with neutral coloring

scikit-learn Integration

SWIFTMonitor inherits from BaseEstimator and TransformerMixin:

# Parameter inspection and modification
monitor.get_params()
monitor.set_params(alpha=0.01, correction="bonferroni")

# fit_transform shorthand
X_transformed = monitor.fit_transform(X_ref)

# Clone with same parameters
from sklearn.base import clone
monitor2 = clone(monitor)

Sample vs. Sample Comparison

By default score() and test() compare monitoring data against the fitted reference. Pass X_compare to compare two arbitrary samples instead — the SHAP transformation (buckets + mean SHAP) is always the one learned from the reference:

# Score: monitoring vs. drifted
scores = monitor.score(X_mon, X_compare=X_drifted)

# Full test: monitoring vs. drifted
result = monitor.test(X_mon, X_compare=X_drifted)
print(result.drifted_features)

Configuration

Parameter	Default	Description
model	(required)	Trained LightGBM or XGBoost model
order	1	Wasserstein order: 1 (W1) or 2 (W2, more sensitive to variance)
n_permutations	1000	Number of permutations for p-value estimation
alpha	0.05	Significance level for MTC
correction	"benjamini-hochberg"	MTC method. Aliases: "bh", "fdr", "bonf"
n_synthetic	10	Synthetic observations for empty buckets during fit
max_samples	None	Max pool size for permutation test (subsamples if exceeded)
random_state	42	RNG seed for reproducibility

Development

# Install dev dependencies
uv pip install -e ".[dev]"

# Run tests
pytest tests/

# Run tests with coverage
pytest tests/ --cov=swift --cov-report=term-missing

# Lint
ruff check src/ tests/

Project Structure

swift/
├── src/swift/
│   ├── __init__.py          # Public API exports
│   ├── pipeline.py          # SWIFTMonitor (main entry point)
│   ├── plotting.py          # Visualization functions
│   ├── extraction.py        # Stage 1: Decision point extraction
│   ├── bucketing.py         # Stage 2: Bucket construction
│   ├── normalization.py     # Stage 3: SHAP normalization
│   ├── distance.py          # Stage 4: Wasserstein distance
│   ├── threshold.py         # Stage 5: Permutation test + MTC
│   ├── aggregation.py       # Model-level score aggregation
│   └── types.py             # Data types (BucketSet, SWIFTResult, etc.)
├── tests/                   # Test suite
├── experiments/             # Experiment runners (ablations, power analysis, etc.)
├── notebooks/
│   └── getting_started.ipynb
└── pyproject.toml

Citation

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