Session 8: Aggressive dependency cut → 4.0.0a1

Baseline: end of session 7 (4.0.0a0 tag pushed, GitHub Release published, user testing on Google Colab surfaced "deps still too heavy"). Environment: unchanged.

Theme: owner asked for another round of dep discipline — "lets start cleaning them up. we dont need all kind of tuners. we dont need all that kind of extra visualization... nobody uses all those long tail functionalities. we dont need kmodes, lightgbm, catboost, xgboost, let user install it separately and add them to the library model container or something." Plus: unpin sklearn.

BREAKING — extras collapsed#

• REMOVED, BREAKING — pycaret[models] extra entirely removed. xgboost, catboost, kmodes, mlxtend, lightgbm — all no longer installed by pycaret[full] or any other extra. Users install them directly: pip install xgboost lightgbm catboost. Pycaret's model containers already use _check_soft_dependencies to auto-detect and light up when the backing package is present (LGBM container gained this guard in this session; xgboost/catboost already had it).
• REMOVED, BREAKING — pycaret[tuners] extra entirely removed. optuna, optuna-integration, scikit-optimize, hyperopt — all gone. Users attach a custom search via sklearn's GridSearchCV / RandomizedSearchCV / HalvingRandomSearchCV, or install an optuna-type backend and pass a constructed search-cv to tune_model.
• REMOVED, BREAKING — pycaret[analysis] extra entirely removed. shap, interpret, umap-learn — all gone. Interpretability re-enters targeted in a later release (owner: "for analysis we get drop interpret related functionalities, shap as well for now, we will bring it later").
• REMOVED, BREAKING — pycaret[prophet] extra entirely removed.

DEPS — core cut from 19 → 11#

Removed from core dependencies:

• lightgbm>=4.3 (moved to user-install).
• cloudpickle>=3.0 (joblib pulls it transitively when needed; load_experiment() lazy-imports with a clean fallback).
• psutil>=5.9 (system-info logging falls back to os.cpu_count()).
• xxhash>=3.4 (FastMemory falls back to hashlib.blake2b(digest_size=16)).
• matplotlib>=3.9 (Plotly is the single chosen library; matplotlib lazy-imported in 4 residual non-Plotly call sites).
• kaleido>=0.2 (moved to new export extra).
• nbformat>=5.10 (moved to notebook extra).
• ipywidgets>=8.1 (moved to notebook extra).

Kept in core:

• numpy>=1.26, pandas>=2.2, scipy>=1.11, scikit-learn>=1.7 (upper cap removed — see below), joblib>=1.4, plotly>=5.22, tqdm>=4.66, requests>=2.32, jinja2>=3.1 (pandas.style), ipython>=8.18.

Transitional (still in core, flagged for removal):

• imbalanced-learn>=0.13 — removed when pycaret.internal.pipeline.Pipeline stops inheriting from imblearn.pipeline.Pipeline (Phase 4).
• category-encoders>=2.6 — removed when legacy preprocessor is rewritten on native sklearn encoders (Phase 4).

CHANGED — sklearn unpinned#

• CHANGED — scikit-learn>=1.7 — upper cap <1.8 removed. The cap existed only because sktime (in the timeseries extra) required scikit-learn<1.8. Since sktime is no longer pulled into the default install — only when a user installs pycaret[timeseries] — the core can track the latest sklearn. Default fresh install now pulls sklearn 1.8, NumPy 2.4, pandas 3.0.

REMOVED — show_versions diagnostic table entries#

• CHANGED — pycaret/utils/_show_versions.py dependency-version reporting table no longer includes mlxtend, kmodes, kaleido, jinja2 (still reported), xxhash (uses stdlib now). Kept for diagnostic: numpy, pandas, scipy, scikit-learn, joblib, plotly, tqdm, requests, ipython, imbalanced-learn, category-encoders, matplotlib (if installed), plus extras packages when detected.

FIXED — soft-dep guards#

• FIXED — pycaret/containers/models/classification.py::LGBMClassifierContainer — added _check_soft_dependencies("lightgbm") guard; container sets self.active = False cleanly when lightgbm isn't installed. Same guard added to regression.py::LGBMRegressorContainer and time_series.py::LGBMCdsDtContainer (TS variant also pre-sets is_gpu_enabled = False so the god-class __init__ flow doesn't AttributeError before the active=False short-circuit takes effect).
• FIXED — pycaret/internal/memory.py — xxhash import now try/except; falls back to hashlib.blake2b. No behaviour change in practice (hash-based cache key collisions don't matter for correctness, only perf).
• FIXED — pycaret/internal/pycaret_experiment/pycaret_experiment.py — cloudpickle now lazy-imported; psutil system-info logging is conditional.
• FIXED — pycaret/internal/pycaret_experiment/supervised_experiment.py — matplotlib.pyplot lazy-imported. Only the two plt.savefig() interpret-model call sites need it; both guarded against plt is None via the surrounding if self.logging_param block and the try/except AttributeError in the calling code.
• FIXED — pycaret/internal/plots/helper.py + pycaret/internal/plots/utils/time_series.py — matplotlib lazy-imported; MatplotlibDefaultDPI already had try/except wrapping its plt.rcParams access so plt = None is a silent no-op.

TESTS#

• TESTS — 32/32 green on the lean 4.0.0a1 in 71s on the full-extras dev venv.
• TESTS — Fresh pycaret install with no extras (41 deps total) successfully runs ClassificationExperiment(target='Purchase').fit(juice) → compare_models(include=['lr','dt']) → predict_model → save_model → load_model roundtrip on sklearn 1.8 / NumPy 2.4 / pandas 3.0.

BUILD#

• BUILD, BREAKING — Version 4.0.0a0 → 4.0.0a1.
• BUILD — Rebuilt wheel pycaret-4.0.0a1-py3-none-any.whl (412 KB, 112 files). twine check PASSED.

DOCS#

• DOCS — CHANGELOG.md — prepended 4.0.0a1 release entry with install commands, changed-removed-fixed subsections, new-extras-structure table, and the two "transitional deps" flagged for removal.
• DOCS — this session-8 block in release_notes_pycaret4.md.
• DOCS — KILL_LIST.md updated with the 4.0.0a1 extras collapse.

Session 8 delta summary#