shaply

Usual SHAP explainability figures, rendered as interactive Plotly charts.

shaply reproduces the familiar figures from the shap library - bar, beeswarm, waterfall, dependence (scatter) and heatmap - but returns plotly.graph_objects.Figure objects instead of matplotlib axes, so the plots are interactive and embeddable out of the box.

It does not depend on shap: every plotting function accepts a shap.Explanation-like object, a raw NumPy array of SHAP values, or a pandas DataFrame.

Install

uv add shaply
# or
pip install shaply

Optional dependencies (extras)

shaply itself only needs numpy, plotly and pydantic. Extra features and the example notebook pull in heavier packages, grouped as installable extras:

uv add "shaply[pandas]"       # DataFrame support
uv add "shaply[examples]"     # run the demo notebook
# or with pip
pip install "shaply[examples]"

Building the wheel from source (use the package without PyPI)

To build and install shaply straight from a clone of this repository, without going through PyPI:

git clone https://github.com/antoine126/shaply.git
cd shaply
uv build --wheel     # produces dist/shaply-<version>-py3-none-any.whl

Then install the wheel wherever you need it:

uv add /path/to/shaply/dist/shaply-<version>-py3-none-any.whl
# or with pip, in any environment
pip install /path/to/shaply/dist/shaply-<version>-py3-none-any.whl

The wheel is self-contained (it ships py.typed, so type annotations reach the installed package) and does not require uv or the source tree at runtime.

Quick start

import shaply

# `explanation` can be a shap.Explanation, an ndarray, or a DataFrame
fig = shaply.beeswarm(explanation)
fig.show()

fig = shaply.bar(explanation)
fig = shaply.waterfall(explanation, sample_index=0)
fig = shaply.scatter(explanation, feature="income", color_feature="age")
fig = shaply.heatmap(explanation)

Every function takes an optional typed config from shaply.config:

from shaply.config import BeeswarmConfig
from shaply.enums import ColorScale, FeatureOrdering

cfg = BeeswarmConfig(
    max_display=15,
    ordering=FeatureOrdering.IMPORTANCE,
    color_scale=ColorScale.RED_BLUE,
)
fig = shaply.beeswarm(explanation, config=cfg)

Available plots

Advanced tools - beyond the usual SHAP plots

These are shaply-only figures aimed at engineers and business-facing data scientists who want to act on SHAP, not just explain a model. They cross SHAP values with the real data to surface operating ranges, tipping points, coupled effects and failure drivers.

Read them as associational, not causal. SHAP measures a feature's contribution to the model's output, not to reality. Wording is deliberately cautious ("associated with", "tipping point of the model") - a strong signal here is a lead to investigate, not a proven cause.

# Beeswarm + real value ranges (needs feature values via data=...)
shaply.beeswarm_ranges(explanation).show()

# Response curve of one feature, with tipping-point detection
shaply.response_curve(explanation, "temperature").show()

# Pairwise interaction strength (needs SHAP *interaction* values)
inter = shap.TreeExplainer(model).shap_interaction_values(X)  # (n, f, f) - pick a class if 4D
shaply.interaction_heatmap(inter, feature_names=list(X.columns)).show()

# What drives the model's mistakes
shaply.error_analysis(explanation, y_true=y_test, y_pred=model.predict(X_test)).show()

# 2D SHAP surface over a feature pair
shaply.shap_surface(explanation, "temperature", "pressure").show()

# Importance split by an operating regime (quantiles of another feature)
shaply.importance_by_cohort(explanation, by_feature="load").show()

# Redundant features (correlated SHAP), and typical decision patterns
shaply.feature_clustering(explanation).show()
shaply.explanation_archetypes(explanation).show()

# Robustness of the ranking, and monotonicity of each effect
shaply.importance_ci(explanation).show()
shaply.monotonicity_check(explanation).show()

Each takes a typed config from shaply.config (e.g. ResponseCurveConfig, ShapSurfaceConfig, ImportanceByCohortConfig, FeatureClusteringConfig, ExplanationArchetypesConfig, ImportanceCIConfig, MonotonicityConfig).

The clustering and statistics behind these tools are implemented in pure NumPy, so the advanced tools add no runtime dependency beyond numpy/plotly/pydantic.

Example notebook

examples/shaply_demo.ipynb is a full, executed walkthrough. It builds a synthetic dataset with make_classification (5 informative, 2 redundant and 3 pure-noise features), trains five very different classifiers - RandomForest, XGBoost, LightGBM, LogisticRegression and an RBF SVM - computes SHAP values for each (TreeExplainer, LinearExplainer, KernelExplainer) and renders every shaply figure for all of them, plus the advanced tools (response curve, interaction heatmap, error analysis) and a cross-model importance comparison.

uv sync --extra examples
uv run jupyter lab examples/shaply_demo.ipynb
# regenerate the executed outputs from scratch:
uv run jupyter nbconvert --to notebook --execute --inplace examples/shaply_demo.ipynb

Development

uv sync
uv run ruff check . --fix
uv run ruff format .
uv run mypy
uv run pytest

The package ships a py.typed marker, so all type annotations are available to downstream users.

License

MIT