SHAP v0.54.1

Highlights

CUDA acceleration with shap-cuda

GPU support now lives in the separate shap-cuda package, providing the CUDA-accelerated backend for shap.GPUTreeExplainer. The GPU wheel workflow builds CUDA-enabled wheels and tests them on AWS GPU runners before publishing. By @CloseChoice in #5140.

The GPU TreeExplainer notebook has also been refreshed by @RAMitchell in #5220.

What's Changed

Enhancements

• shap.plots.violin now accepts an ax argument and returns the axes, making it easier to customize plots and compose subplots. By @adt-kmr in #5218.

Maintenance

• Remove unused global-path-dependent Tree SHAP code. By @RAMitchell in #5223.

Documentation

• Fix typos in the "An introduction to explainable AI" notebook. By @Yaqiao-chi-helen in #5225.
• Fix typos in comments, docstrings, and an error message. By @guptaishaan in #5213.
• Fix typos and replace the deprecated API reference in the LinearExplainer correlation notebook with maskers.Impute(data). By @ojuschugh1 in #5172.

New Contributors

• @Yaqiao-chi-helen made their first contribution in #5225.
• @guptaishaan made their first contribution in #5213.
• @ojuschugh1 made their first contribution in #5172.

Full Changelog: v0.53.0...v0.54.1

Full Changelog: v0.53.0...v0.54.1

Highlights

• Python 3.15 support: SHAP now builds wheels and runs tests against Python 3.15, including pre-releases. By @CloseChoice in #5143
• Migration away from Numba and llvmlite: core utilities have been migrated to nanobind/C++, allowing Numba and llvmlite to be removed as required dependencies and enabling support for free-threaded Python. By @daidahao and @CloseChoice in #5081, #5082, #5088, #5089, #5090, #5091, #5092, #5117, #5129
• GPU TreeSHAP improvements: LightGBM categorical GPU TreeSHAP support has been added, alongside TreeExplainer changes and CPU/GPU equivalence testing for categorical features. By @RAMitchell and @CloseChoice in #5020 and #5026
• Additional build targets: add Windows ARM64 and Emscripten builds. By @ndabas and @CloseChoice in #5137 and #5162

What's Changed

Breaking changes

• Match KernelExplainer DataFrame columns to the background by name. Duplicated column names now raise an error, but worked previously. By @Atishyy27 in #5164.

Fixes

• Fix int16 overflow of node indices in independent TreeSHAP, which could cause segmentation faults for trees with more than 32,767 nodes. By @clark874 in #5163
• Refactor short circuiting of tree_path_dependent SHAP calculation by @CloseChoice in #5069
• Remove outdated TensorFlow version warnings from DeepExplainer by @SushantTusharJoshi in #5126

Enhancements

• Support LightGBM categorical GPU TreeSHAP by @RAMitchell in #5020
• Use TreeEnsemble directly in TreeExplainer and add CPU/GPU equivalence tests for categorical features by @CloseChoice in #5026
• Add GPU build, packaging, CI, smoke tests, and TestPyPI support by @CloseChoice in #5057, #5058, #5114, #5123, #5127
• Move optional dependencies into dependency groups by @daidahao in #4589
• Add performance monitoring for explainers by @daidahao in #5071
• Remove Numba from the image masker by @daidahao in #5089
• Migrate clustering utilities to nanobind by @daidahao in #5088
• Migrate tabular masking utilities to nanobind by @daidahao in #5090
• Migrate ExactExplainer interaction utility to nanobind by @daidahao in #5091
• Migrate lower_credit to nanobind by @daidahao in #5081
• Migrate mask construction to nanobind by @daidahao in #5082
• Migrate masked-model output utilities to nanobind by @daidahao in #5092
• Remove Numba and llvmlite from dependencies and support free-threaded Python by @CloseChoice in #5129
• Run tests and build wheels for Python 3.15, including pre-releases, by @CloseChoice in #5143
• Add Windows ARM64 builds by @ndabas in #5137
• Add Emscripten builds by @CloseChoice in #5162

Other Changes

• Remove AppVeyor and .claude files from the build directory by @CloseChoice in #5023
• Increase test coverage for shap/maskers/_image.py by @bedupako12mas in #4519
• Document ASV performance monitoring by @daidahao in #5083
• Remove llvmlite and Numba restrictions on macOS by @CloseChoice in #5117
• Add type hints for plots._benchmark by @siddharth7113 in #5101
• Build and upload nightly wheels to scientific-python-nightly-wheels by @CloseChoice in #5136
• Skip nightly wheel builds when the default branch is unchanged by @CloseChoice in #5141
• Update actions/checkout to v7.0.1 and pin its commit SHA by @CloseChoice in #5144
• Update Dependabot configuration by @CloseChoice in #5145
• Rename master to main by @CloseChoice in #5153
• Add docstrings to exception classes in _exceptions.py by @adt-kmr in #5208

New Contributors

• @bedupako12mas made their first contribution in #4519
• @siddharth7113 made their first contribution in #5101
• @SushantTusharJoshi made their first contribution in #5126
• @ndabas made their first contribution in #5137
• @Atishyy27 made their first contribution in #5164
• @clark874 made their first contribution in #5163
• @adt-kmr made their first contribution in #5208

Full Changelog: v0.52.0...v0.53.0

Highlights

• Python 3.15.0rc2 support: SHAP now builds and runs tests against Python 3.15 pre-releases, including Python 3.15.0rc2. By @CloseChoice in #5143
• Migration away from Numba and llvmlite: core utilities have been migrated to nanobind/C++, allowing Numba and llvmlite to be removed as required dependencies and enabling support for free-threaded Python. By @daidahao and @CloseChoice in #5081, #5082, #5088, #5089, #5090, #5091, #5092, #5117, #5129
• GPU TreeSHAP improvements: LightGBM categorical GPU TreeSHAP support has been added, alongside TreeExplainer changes and CPU/GPU equivalence testing for categorical features. By @RAMitchell and @CloseChoice in #5020 and #5026

What's Changed

Fixes

• Refactor short circuiting of tree_path_dependent shap calculation by @CloseChoice in #5069
• Remove outdated TensorFlow version warnings from DeepExplainer by @SushantTusharJoshi in #5126

Enhancements

• Support LightGBM categorical GPU TreeSHAP by @RAMitchell in #5020
• Use TreeEnsemble directly in TreeExplainer and add cpu gpu equivalence tests for categories by @CloseChoice in #5026
• Add GPU build, packaging, CI, smoke tests, and TestPyPI support by @CloseChoice in #5057, #5058, #5114, #5123, #5127
• ENH: Move optional-dep into dep-groups by @daidahao in #4589
• ENH: Performance monitoring for explainers by @daidahao in #5071
• PERF: Remove Numba from image masker by @daidahao in #5089
• PERF: Migrate clustering utilities to nanobind by @daidahao in #5088
• PERF: Migrate tabular masking utilities to nanobind by @daidahao in #5090
• PERF: Migrate ExactExplainer interaction utility to nanobind by @daidahao in #5091
• PERF: Migrate lower_credit to nanobind by @daidahao in #5081
• PERF: Migrate mask construction to nanobind by @daidahao in #5082
• PERF: Migrate masked-model output utilities to nanobind by @daidahao in #5092
• remove numba and llvmlite from dependencies, support free-threaded python by @CloseChoice in #5129
• run tests and build wheels for python 3.15 (including pre-release) by @CloseChoice in #5143
• Add Windows ARM64 build by @ndabas in #5137

Other Changes

• TST: Increase test coverage for shap/maskers/_image.py by @bedupako12mas in #4519
• DOC: Document ASV performance monitoring by @daidahao in #5083
• remove llvmlite and numba restrictions on macos by @CloseChoice in #5117
• TYP: add type hints for plots._benchmark by @siddharth7113 in #5101
• Build and upload nightly wheels to scientific-python-nightly-wheels by @CloseChoice in #5136
• update bundle js by @CloseChoice in #5139
• Skip nightly wheel build when master is unchanged by @CloseChoice in #5141
• Update actions/checkout to v7.0.1 and pin its commit SHA by @CloseChoice in #5144
• update dependabot by @CloseChoice in #5145

New Contributors

• @bedupako12mas made their first contribution in #4519
• @siddharth7113 made their first contribution in #5101
• @SushantTusharJoshi made their first contribution in #5126
• @ndabas made their first contribution in #5137

Full Changelog: v0.52.0...v0.53.0rc0

What's Changed

Highlights

• Native bindings rewritten with nanobind, and the build system migrated
  from setup.py to scikit-build-core + CMake. The Cython
  _kernel_lib.pyx and the existing _cext / _cext_gpu modules are now
  built as nanobind extensions, and MANIFEST.in / setup.py have been
  removed in favor of a pyproject.toml-driven build. (#4366, by
  @CloseChoice and @daidahao and @claude)
• Minimum dependency versions raised to follow SPEC 0. (#4310, by @CloseChoice)

Enhancements

• Log an INFO message when background data is sub-sampled (closes #3461).
  (#4294, by @Ebube22)
• Add cross-links between the API Reference and API Examples pages.
  (#4329, by @Abhishek9639)
• Drop the unused typing-extensions dependency. (#4629, by @samudraneel05)

Bug Fixes

• Improve GPU TreeExplainer parity: preserve XGBoost default/missing child
  routing so NaN values follow the same branch as the source model, preserve
  vector-valued XGBoost base_score values (fixing multiclass additivity
  offsets), and tighten categorical handling so sklearn models with
  enable_categorical=True raise the existing unsupported-categorical error
  even when the booster does not expose feature_types. (#4997, by
  @RAMitchell)
• TreeExplainer no longer crashes with pandas nullable dtypes.
  (#4298, by @tudstudent)
• Fix NameError when with_binary=False and with_cuda=True by
  initializing compile_args. (#4322, by @mohityadav8)
• Text plot colors render correctly by converting NumPy types to float.
  (#4332, by @Saloni-0465)
• Assign the result of np.flipud so it is no longer a no-op in
  AdditiveForceArrayVisualizer. (#4343, by @Mahaveerjain-18)
• Use isinstance() instead of is for type checks. (#4373, by @divyam-jha123)
• Record 'sample' in Explanation.op_history instead of '__getitem__'.
  (#4376, by @tarun-227)
• Move the unreachable large-dataset warning into the interventional branch.
  (#4392, by @Mahaveerjain-18)
• Fix "structure of inputs doesn't match the expected structure".
  (#4410, by @shaivimalik)
• Fix the return value from is_color_map() in plots/_beeswarm.py.
  (#4421, by @maniktyagi04)
• Suppress the UserWarning when LGBMRegressor was fitted with feature
  names. (#4422, by @shaivimalik)
• Correct max_evals calculation in PermutationExplainer.shap_values().
  (#4434, by @AftAb-25)
• Fix incorrect measure calls in benchmark/metrics.py.
  (#4447, by @Jiya873)
• Handle zero division in the heatmap when all SHAP values are zero.
  (#4449, by @BaibhavKundu2005)
• Correct file path handling in make_dir (image.py).
  (#4501, by @Abhishek-Kumar-Rai5)
• Use isinstance() instead of type() is for tuple checks.
  (#4508, by @neha222222)
• Warn when the covariance matrix is singular in LinearExplainer.
  (#4471, by @BaibhavKundu2005)
• Remove a debug print and a typo in serializable.
  (#4336, by @dhiraj-143r)

Maintenance

• Fix a NumPy deprecation warning in the coalition explainer.
  (#4291, by @CloseChoice)
• Fix deprecated NumPy dtype aliases in test_kernel.py.
  (#4302, by @peteroyce)
• Replace the deprecated input_shape kwarg with an Input layer in tests.
  (#4301, by @kvr06-ai)
• Replace deprecated return_all_scores with top_k=None (#3974).
  (#4393, by @shauryam2807)
• Remove dead code in utils/_keras.py (unused and broken on Keras 3).
  (#4625, by @SuMayaBee)
• Remove the dead __check_cache function and stale cache variables in
  benchmark. (#4485, by @dhiraj-143r)
• Update bundled JS. (#4308, by @CloseChoice)

Type Hints

• Add type hints for explainers.other._random. (#4436, by @SarthakB11)

Tests

• Add SHAP values baselines. (#4307, by @CloseChoice)
• Add more exact explainer
  #4366tests. (#4312, by @CloseChoice)
• Re-enable causalml tests. (#4311, by @dhiraj-143r)
• Make force plot tests deterministic to fix macOS flakiness (#4102).
  (#4315, by @Abhishek9639)
• Add a new exact explainer test. (#4364, by @CloseChoice)
• Use the default matplotlib style for tests. (#4330, by @Anushreebasics)
• Restore pytest discovery for shap.actions. (#4350, by @JRV7903)
• Improve test coverage for the Action base class. (#4514, by @Jiya873)
• Add tests in tests/explainers/test_linear.py. (#4416, by @naxatra2)
• Increase test coverage for shap/explainers/other/_treegain.py.
  (#4427, by @rupeshca007)

Documentation

• Update LightGBM links. (#4296, by @jameslamb)
• Fix Python installation version from 3.11 to 3.12. (#4319, by @kronzter)
• Fix a broken link. (#4400, by @Jiya873)
• Update the Iris classification notebook to use the modern SHAP API.
  (#4318, by @JRV7903)
• Fix typos in the heatmap notebook. (#4359, by @Bot87Ever)
• Reformat the waterfall notebook. (#4378, by @justin212407)
• Improve clarity and documentation in the sentiment analysis notebook.
  (#4397, by @24f3004086)
• Improve the embedding plot docstring. (#4438, by @Jiya873)
• Modernize the "Fitting a Linear Simulation with XGBoost" notebook.
  (#4428, by @vibhor-5)
• Fix spelling errors in source code docstrings and comments.
  (#4446, by @rupeshca007)
• Clarify the expected_value math formula in TreeExplainer (#4414).
  (#4418, by @DashratRajpurohit)

New Contributors

• @daidahao made their first contribution in #4366
• @peteroyce made their first contribution in #4302
• @kvr06-ai made their first contribution in #4301
• @dhiraj-143r made their first contribution in #4311
• @Abhishek9639 made their first contribution in #4315
• @kronzter made their first contribution in #4319
• @Jiya873 made their first contribution in #4400
• @shauryam2807 made their first contribution in #4393
• @Ebube22 made their first contribution in #4294
• @JRV7903 made their first contribution in #4318
• @mohityadav8 made their first contribution in #4322
• @Saloni-0465 made their first contribution in #4332
• @Anushreebasics made their first contribution in #4330
• @Bot87Ever made their first contribution in #4359
• @SuMayaBee made their first contribution in #4625
• @divyam-jha123 made their first contribution in #4373
• @justin212407 made their first contribution in #4378
• @samudraneel05 made their first contribution in #4629
• @Mahaveerjain-18 made their first contribution in #4392
• @24f3004086 made their first contribution in #4397
• @shaivimalik made their first contribution in #4410
• @tarun-227 made their first contribution in #4376
• @maniktyagi04 made their first contribution in #4421
• @naxatra2 made their first contribution in #4416
• @rupeshca007 made their first contribution in #4427
• @vibhor-5 made their first contribution in #4428
• @SarthakB11 made their first contribution in #4436
• @AftAb-25 made their first contribution in #4434
• @DashratRajpurohit made their first contribution in #4418
• @Abhishek-Kumar-Rai5 made their first contribution in #4501
• @BaibhavKundu2005 made their first contribution in #4449
• @neha222222 made their first contribution in #4508

Full Changelog: v0.51.0...v0.52.0

What's Changed

Fixes

• fix: check first that the feature is not in the leaf node by @Far-naz in #4268
• Fix missing array to scalar conversion in MAPLE by @Scienfitz in #4285
• fix Python Version of Tree SHAP notebook by @CloseChoice in #4289
• Fix path-dependent SHAP NaN with small background datasets by @tudstudent in #4272
• Fix IndexError in format_value() when handling empty strings by @Copilot in #4238
• Fix test_scatter_categorical for pandas 3.0 compatibility by @Copilot in #4253
• Fix SamplingExplainer.explain for Series. by @ljw20180420 in #4200

Other Changes

• Add tests for fixed masker by @CloseChoice in #4216
• Add comprehensive type hints to production code by @CloseChoice in #4217
• unpin versions and pin numba, skip causalml test by @CloseChoice in #4235
• Clarify TreeExplainer binary classification returns different shapes per model by @Copilot in #4254
• Add tests for explainers by @CloseChoice in #4218
• Add Colab Notebook to test GPUTreeExplainer by @CloseChoice in #4266
• Fix spelling error in docs: perterbation to perturbation by @laffertyryan0 in #4275
• update test_scatter to work with latest xgboost by @CloseChoice in #4278
• add ai usage policy at appropriate place in contributing, mention contributing in readme by @CloseChoice in #4279
• deprecate testing of newer llvmlite versions on macos x64_86 by @CloseChoice in #4286

New Contributors

• @ljw20180420 made their first contribution in #4200
• @Copilot made their first contribution in #4238
• @laffertyryan0 made their first contribution in #4275
• @tudstudent made their first contribution in #4272
• @Far-naz made their first contribution in #4268
• @Scienfitz made their first contribution in #4285

Full Changelog: v0.50.0...v0.51.0

What's Changed

• hand over threshold_types to GPUTreeExplainer by @CloseChoice in #4181
• Improve base_score assignment by @lsdxp in #4187
• test against python 3.14, remove support for python 3.9 and 3.10 by @CloseChoice in #4176
• Always force transformers label2id ids to integers by @evamaxfield in #4192
• Fix gpu tree explainer tests by @CloseChoice in #4199

New Contributors

• @lsdxp made their first contribution in #4187
• @evamaxfield made their first contribution in #4192

Full Changelog: v0.49.1...v0.50.0

What's Changed

Fix broken v0.49.0 release.

The previous Release wasn't properly published due to HTTP errors on MacOS.

What's Changed

NOTE: this is the last release supporting Python 3.9 AND Python 3.10. From v0.50.0 onwards, we'll only support Python 3.11 and above.

• Add support for categorical splits in C++ library by @nqviller in #4171
• Add options to customize the color bar of the violin plot by @FanwangM in #4119

Other Changes

• plots.image show labels for every row if provided by @julesvanrie in #4113
• use nbtest by @CloseChoice in #4143
• Added Support for PyTorch Flaten by @RoyiAvital in #4148
• Update .save() and .load() methods to remove AttributeError in Explainer and Serializer by @oscarl77 in #4155
• fix for missing objective error for LGBMRegressor by @imatiach-msft and @CloseChoice in #1063
• Fix potential TypeError in numeric feature branch by adding string conversion for feature names #4150 by @YunyuG in #4159
• Improve Coalition Explainer User Guidance and Fix Tree Building by @EnzoFanAccount in #4116

New Contributors

• @diego-pm made their first contribution in #4110
• @julesvanrie made their first contribution in #4113
• @Helias made their first contribution in #4141
• @oscarl77 made their first contribution in #4155
• @FanwangM made their first contribution in #4119
• @YunyuG made their first contribution in #4159
• @EnzoFanAccount made their first contribution in #4116
• @nqviller made their first contribution in #4171

Full Changelog: v0.48.0...v0.49.0

What's Changed

Added

• Add CoalitionExplainer and add possibility of using Winter Values in Partition Explainer by @CousinThrockmorton in #3666
• Support and test against Python 3.13 by @connortann in #3861 and @CloseChoice in #4104
• Add Support for PyTorch Identity Layer by @RoyiAvital in #4028

Documentation

• Update Explaining a model that uses standardized features by @randombenj in #3903

Other Changes

• Changed alias by @Ja-Tink in #4049
• update javascript packages by @CloseChoice in #4067
• Fixed visual bug for small SHAP values by @Ja-Tink in #4053
• Fix summary_plot displays only one feature 4081 by @CloseChoice in #4087
• resolve threading warnings of regex library by @emmanuel-ferdman in #4084

New Contributors

• @RoyiAvital made their first contribution in #4028
• @Ja-Tink made their first contribution in #4049
• @emmanuel-ferdman made their first contribution in #4084
• @randombenj made their first contribution in #3903
• @CousinThrockmorton made their first contribution in #3666

Full Changelog: v0.47.2...v0.48.0

What's Changed

Added

• Add experimental causalml support by @alexander-pv in #3273

Other Changes

• DOCS: clarity for partition tree explanation in Simple California Demo.ipynb by @ethanknights in #4027
• FIX: unique value jitter by @fabianliebig in #4041
• FIX: fix neutral language in SHAP value description by @Hrafz in #4058
• FIX regression test for javascript plotting compontents and add tests by @CloseChoice in #4060

New Contributors

• @ethanknights made their first contribution in #4027
• @Hrafz made their first contribution in #4058

Full Changelog: v0.47.1...v0.47.2