Modernize for scikit-learn 1.6+, NumPy 2.0+, scipy 1.17+, matplotlib 3.7+

[jg585Username]

Summary

Fixes 337 test failures caused by breaking API changes introduced in recent versions of scikit-learn, NumPy, scipy, and matplotlib. The project now supports:

• scikit-learn ≥ 1.6 (tested against 1.8.0)
• NumPy ≥ 1.24 (tested against 2.4.4)
• scipy ≥ 1.7 (tested against 1.17.1)
• matplotlib ≥ 3.5 (tested against 3.10.9)
• Python 3.9 – 3.13

Result: 1127 passed, 0 failed (was 337 failures before this PR).

---

scikit-learn compatibility

_estimator_type removed (sklearn 1.6+)

yellowbrick/utils/types.py — The most widespread breakage. sklearn 1.6 removed the _estimator_type class attribute from many estimators. Added _get_estimator_type() which falls back through: legacy attribute → __sklearn_tags__() API → Mixin subclass inspection. Rewrote is_classifier(), is_regressor(), is_clusterer() on top of it.

Pipeline.__sklearn_is_fitted__ (sklearn 1.8+)

yellowbrick/base.py — sklearn 1.8's Pipeline.score() calls check_is_fitted(last_step) on the final pipeline step. ModelVisualizer has no trailing-underscore attributes of its own, so this raised NotFittedError even after a successful fit(). Added __sklearn_is_fitted__() delegating to check_is_fitted(self.estimator).

__sklearn_tags__() for ContribEstimator (sklearn 1.6+)

yellowbrick/contrib/wrapper.py — Added __sklearn_tags__() so wrapped contrib estimators correctly report their type through the new Tags API.

_check_targets() returns 4-tuple (sklearn 1.8)

yellowbrick/classifier/class_prediction_error.py — Was unpacked as 3-tuple; switched to indexed access.

multi_class="auto" removed (sklearn 1.7)

yellowbrick/classifier/threshold.py — Removed the parameter.

store_cv_values / cv_values_ renamed (sklearn 1.7)

yellowbrick/regressor/alphas.py — Added dual-check with fallback to old names for backward compatibility.

np.matrix rejected (sklearn 1.7)

yellowbrick/cluster/elbow.py — Sparse matrix .mean() returns np.matrix; wrapped with np.asarray().

TSNE perplexity validation (sklearn 1.7+)

yellowbrick/cluster/icdm.py — perplexity must be < n_samples; added dynamic cap before fit_transform.

---

NumPy compatibility

File	Change	Reason
yellowbrick/utils/helpers.py	np.in1d → np.isin	Removed in NumPy 2.0
yellowbrick/contrib/missing/dispersion.py	np.string_ → np.bytes_, np.unicode_ → np.str_	Removed in NumPy 1.24
yellowbrick/text/dispersion.py	Wrapped generators in list() for np.stack()	NumPy 2.0 requires sequences
yellowbrick/cluster/icdm.py	interpolation= → method= in np.percentile	Renamed in NumPy 1.22
yellowbrick/classifier/base.py	Extract KeyError key via e.args[0]	NumPy 2.0 changed np.str_ repr

---

matplotlib compatibility

yellowbrick/regressor/influence.py — Removed use_line_collection=True from ax.stem() (argument removed in matplotlib 3.7).

---

Test suite updates

• 13 test files updated for deprecated APIs, wrong MRO order, removed parameters, and updated numerical expected values (sklearn 1.8 / scipy 1.17 produce slightly different scores)
• ~96 baseline PNG images regenerated to match matplotlib 3.10 rendering
• tests/test_compat.py — 25 new targeted regression tests, one per fix, so future package upgrades immediately surface regressions:
  • TestEstimatorTypeDetection — verifies is_classifier/regressor/clusterer via Mixin-only classes (no _estimator_type)
  • TestModelVisualizerFittedState — verifies __sklearn_is_fitted__ and Pipeline.score() don't raise after fit
  • TestNumpyCompat — documents and verifies all NumPy 2.0 API removals
  • TestMatplotlibCompat — verifies ax.stem() without use_line_collection
• requirements.txt / setup.py — bumped minimums and added Python 3.9–3.13 classifiers

[jg585Username]

Relationship to Other Open Modernization PRs

While working on this PR, I noticed several other open attempts at the same problem. Here's how they relate:

#1322 and #1325 (lwgray) — Both are subsets of this PR. #1322 only removes the use_line_collection parameter from ax.stem(). #1325 covers that plus a handful of other matplotlib/NumPy fixes, but stops well short of full compatibility.

#1332 and #1333 (PythonCharmers) — These fix types.py and contrib/wrapper.py only, targeting the sklearn 1.6 tags API. Their approach for types.py (delegating directly to sklearn's own is_classifier/regressor/clusterer) is arguably simpler than ours, but neither PR addresses the sklearn 1.8 Pipeline.__sklearn_is_fitted__ breakage or any of the NumPy/matplotlib issues.

#1329 (lwgray) — The most comprehensive existing attempt, targeting sklearn 1.7, NumPy 2.0, and matplotlib 3.10. Covers roughly 70% of the same ground as this PR. The key differences are:

• They skip the pipeline validation tests rather than fixing the root cause; this PR adds __sklearn_is_fitted__() to ModelVisualizer so Pipeline.score() works correctly on sklearn 1.8
• They don't address _check_targets() returning a 4-tuple (sklearn 1.8), TSNE perplexity validation, np.matrix rejection, or multi_class="auto" removal
• This PR targets sklearn 1.8 specifically and achieves 1127 passed, 0 failed
• This PR adds tests/test_compat.py with 25 targeted regression tests so future package upgrades surface breakages immediately

In short: this PR is a strict superset of all the above. If the maintainers prefer a smaller, more incremental approach, merging #1329 first and then layering the remaining sklearn 1.8 fixes on top would also be a viable path.

[lwgray]

@jg585Username I will take a look a this this weekend