sklplus

中文文档 · 中文项目页 · Wiki

sklplus 0.2.0 is a thin umbrella around scikit-learn-style tabular estimators. Most symbols are re-exports (often the same object as upstream). A few adapters and custom preprocessors fill gaps we care about when building Pipelines by hand.

It is for people who want from sklplus.… import … and standard fit / transform / predict composition. It is not AutoML: there is no get_model, no setup(), and no string-ID model factory.

Fits: classification, regression, clustering, and anomaly detection on tabular data, with preprocess / sampling / selection / search / metrics helpers re-exported under sklearn-like paths.

Does not fit (v0.1): time-series or NLP specialist modules. If you need PyCaret-style experiment sessions, use something else.

Requirements

• Python >=3.10
• Default install: scikit-learn>=1.3, numpy, pandas, scipy, joblib
• Optional extras:
  • [boost]: xgboost, lightgbm, catboost
  • [imblearn]: imbalanced-learn
  • [anomaly]: pyod
  • [extra]: kmodes, category-encoders, feature-engine
  • [all]: union of the above
  • [dev]: pytest>=7, ruff

Install

PyPI / import / GitHub repo name: sklplus (https://github.com/IncubatorShokuhou/sklplus).

pip install sklplus
# with boosting libs
pip install "sklplus[boost]"
# everything formerly bundled by default
pip install "sklplus[all]"

From a clone (editable, with test tools and full extras):

pip install -e ".[dev,all]"

Missing optional packages raise ImportError naming the extra (e.g. pip install "sklplus[boost]").

Minimal usage

from sklplus.linear_model import LogisticRegression, Ridge
from sklplus.ensemble import RandomForestClassifier, XGBClassifier
from sklplus.preprocessing import (
    StandardScaler,
    CleanColumnNames,
    RareCategoryGrouper,
    DateFeatureExtractor,
    TargetLabelEncoder,
)
from sklplus.sampling import SMOTE
from sklplus.anomaly import IForest
from sklplus.pipeline import Pipeline, ImbPipeline
from sklplus.compose import ColumnTransformer

Re-exports aim to be identity with upstream when possible, e.g. sklplus.ensemble.RandomForestClassifier is sklearn.ensemble.RandomForestClassifier.

Boosting import paths

XGBClassifier / LGBMClassifier / CatBoostClassifier (and the regressors) are exported from both sklplus.ensemble and sklplus.xgboost / lightgbm / catboost. They are the same class object:

from sklplus.ensemble import XGBClassifier as A
from sklplus.xgboost import XGBClassifier as B
assert A is B

Task hubs and ensemble technique slices

sklearn-style modules (linear_model, ensemble, xgboost, …) stay put. Two extra hubs re-export the same objects by task:

from sklplus.classification import LogisticRegression, XGBClassifier
from sklplus.regression import Ridge, XGBRegressor
from sklplus.linear_model import LogisticRegression as LM
from sklplus.xgboost import XGBClassifier as XGB
assert LogisticRegression is LM
assert XGBClassifier is XGB

sklplus.ensemble is still the full umbrella. Technique submodules are slices, not moves:

• sklplus.ensemble.bagging — BaggingClassifier / BaggingRegressor only (RF / ExtraTrees stay on the umbrella)
• sklplus.ensemble.boosting — AdaBoost, GradientBoosting, XGB, LGBM, CatBoost
• sklplus.ensemble.stacking — StackingClassifier / StackingRegressor
• sklplus.ensemble.voting — VotingClassifier / VotingRegressor

from sklplus.ensemble.boosting import XGBClassifier as A
from sklplus.xgboost import XGBClassifier as B
from sklplus.ensemble import XGBClassifier as C
assert A is B is C

There is no top-level sklplus.stacking, and no get_model / factory.

Resampling and pipeline checks

Use ImbPipeline when a step calls fit_resample (e.g. SMOTE, RemoveOutliers).

sklplus.pipeline.Pipeline is a thin subclass of sklearn's Pipeline, so sklplus.pipeline.Pipeline is sklearn.pipeline.Pipeline is False. Construction defaults to check_conflicts=True. make_pipeline builds this wrapper. Disable with check_conflicts=False, or call validate_pipeline_steps(steps, *, kind) directly.

Examples

Scripts under examples/:

python examples/01_classification_pipeline.py

Custom preprocessors in 0.1.2

These live in sklplus.preprocessing (alongside re-exported scalers/encoders):

P0

• CleanColumnNames
• DateFeatureExtractor
• RareCategoryGrouper
• TargetLabelEncoder

P1

• GroupFeatures — row-wise aggregates over named column groups
• RemoveMulticollinearity — drop highly correlated numeric features (optional prefer-by-y)
• IterativeImputerPlus — IterativeImputer for numeric + most-frequent for categoricals
• TextEmbedder — BoW / TF-IDF on text columns, concatenated back

Sampling (not preprocessing): RemoveOutliers in sklplus.sampling is an imblearn-style sampler (fit_resample) for train-only outlier row drops — use with ImbPipeline.

Anomaly detectors under sklplus.anomaly are thin wrappers around pyod (IForest, LOF, …) so they can sit in a sklearn Pipeline.

Known gaps

Still open: Optuna/skopt search wrappers. Custom preprocessors pass sklearn.base.clone and check_estimator with DataFrame-only tags (two_d_array=False, so ndarray suites are skipped by design); TargetLabelEncoder / RemoveOutliers / anomaly adapters have clone smoke. P1 preprocessors / RemoveOutliers, pipeline conflict checks, and classification/regression hubs are in. Optional heavy deps are extras since 0.2.0.

License: MIT (see LICENSE).

Contributing: see CONTRIBUTING.md.