Version history: scikit-learn
v1
scan: pass2026-08-11 — Ingested from sourceInitial version.
v2
currentscan: pass2026-11-02 — Currency pass 2026-11-01: custom estimators use the public validate_data()/check_is_fitted() helpers (scikit-learn 1.6+) instead of private _validate_data; the Cython section built on internal sklearn.utils._typedefs is replaced by built-in performance features (n_jobs, HistGradientBoosting, Pipeline memory); python 3.12; ArrayLike hints; quoted globs so the frontmatter parses.---- description: Definitive guidelines for writing robust, maintainable, and performant scikit-learn code, emphasizing consistent preprocessing, API adherence, and data leakage prevention. - globs: **/*.py+ description: "Guidelines for robust, maintainable scikit-learn code: pipelines everywhere, API-compliant custom estimators, data-leakage prevention, reproducibility, and check_estimator tests." + globs: "**/*.py" + alwaysApply: false --- # scikit-learn Best Practices - This guide outlines our team's definitive best practices for using and extending scikit-learn. Adhering to these rules ensures consistent, reproducible, and production-ready machine learning code.+ This guide outlines best practices for using and extending scikit-learn (1.6+). Adhering to these rules ensures consistent, reproducible, and production-ready machine learning code. ## 1. Code Organization and Structure ### 1.1. Always Use Pipelines for Preprocessing and Models **Pipelines are mandatory.** They prevent data leakage, ensure consistent transformations across training and inference, and simplify hyperparameter tuning. ❌ BAD: Inconsistent manual transformations ```python from sklearn.preprocessing import StandardScaler from sklearn.linear_model import LinearRegression from sklearn.model_selection import train_test_split from sklearn.metrics import mean_squared_error import numpy as np X, y = np.random.rand(100, 5), np.random.rand(100) X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42) scaler = StandardScaler() X_train_scaled = scaler.fit_transform(X_train) model = LinearRegression().fit(X_train_scaled, y_train) # Forgetting to scale X_test leads to incorrect predictions y_pred = model.predict(X_test) print(f"MSE (BAD): {mean_squared_error(y_test, y_pred):.2f}") ``` ✅ GOOD: Encapsulate all steps in a `Pipeline` ```python from sklearn.pipeline import make_pipeline from sklearn.preprocessing import StandardScaler from sklearn.linear_model import LinearRegression from sklearn.model_selection import train_test_split from sklearn.metrics import mean_squared_error import numpy as np X, y = np.random.rand(100, 5), np.random.rand(100) X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42) # All steps are chained and applied consistently model = make_pipeline(StandardScaler(), LinearRegression()) model.fit(X_train, y_train) y_pred = model.predict(X_test) print(f"MSE (GOOD): {mean_squared_error(y_test, y_pred):.2f}") ``` ### 1.2. Custom Estimators Must Adhere to the scikit-learn API When creating custom transformers or models, strictly follow the scikit-learn estimator API for seamless integration with pipelines and model selection tools. - Inherit from `BaseEstimator` and relevant mixins (`TransformerMixin`, `ClassifierMixin`, `RegressorMixin`). - `__init__` stores *only* hyperparameters as keyword arguments with defaults. No logic or data processing here. - `fit(self, X, y=None)` learns parameters from data. Store learned attributes with a trailing underscore (e.g., `self.mean_`). Return `self`. - `transform(self, X)` (for transformers) or `predict(self, X)`/`predict_proba(self, X)` (for predictors) applies the learned transformation/prediction. - Implement `get_params` and `set_params` (usually inherited from `BaseEstimator`).+ - Validate input with the public `sklearn.utils.validation.validate_data(self, X, ...)` (scikit-learn 1.6+; it replaced the private `_validate_data` method) and guard `transform`/`predict` with `check_is_fitted(self)`. ❌ BAD: Non-compliant custom estimator ```python class BadCustomScaler: def __init__(self, scale_factor): # Logic in __init__ if scale_factor <= 0: raise ValueError("Scale factor must be positive") self.scale_factor = scale_factor self.mean_ = None # Should be learned in fit def fit(self, X): self.mean_ = X.mean(axis=0) return self def transform(self, X): if self.mean_ is None: raise RuntimeError("Fit not called yet.") return (X - self.mean_) * self.scale_factor # This won't work with GridSearchCV or Pipelines reliably ``` ✅ GOOD: API-compliant custom estimator ```python from sklearn.base import BaseEstimator, TransformerMixin+ from sklearn.utils.validation import check_is_fitted, validate_data import numpy as np class GoodCustomScaler(BaseEstimator, TransformerMixin): def __init__(self, scale_factor: float = 1.0): # Only store hyperparameters, no logic self.scale_factor = scale_factor def fit(self, X: np.ndarray, y=None):- # Input validation (optional, but good practice) - X = self._validate_data(X)+ # Public validation helper: checks shape/dtype and records n_features_in_ + X = validate_data(self, X) if self.scale_factor <= 0: raise ValueError("Scale factor must be positive") self.mean_ = X.mean(axis=0) # Learned attribute with underscore return self def transform(self, X: np.ndarray) -> np.ndarray:- X = self._validate_data(X) # Ensure consistency - if not hasattr(self, "mean_"): - raise RuntimeError("Estimator not fitted. Call fit() first.")+ check_is_fitted(self) # Raises NotFittedError before fit() + X = validate_data(self, X, reset=False) # Same feature count as fit return (X - self.mean_) * self.scale_factor # This estimator integrates perfectly with scikit-learn utilities. ``` ## 2. Common Patterns and Anti-patterns ### 2.1. Prevent Data Leakage: Split First, Fit on Train Only Always split your data into training and test sets *before* any preprocessing. `fit` and `fit_transform` methods must *only* be called on the training data. `transform` can then be called on both training and test data. ❌ BAD: Data leakage by fitting on all data ```python from sklearn.preprocessing import StandardScaler from sklearn.model_selection import train_test_split import numpy as np X, y = np.random.rand(100, 5), np.random.rand(100) X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42) scaler = StandardScaler() # Fitting on all data (X) leaks information from the test set X_scaled = scaler.fit_transform(X) X_train_scaled = X_scaled[:len(X_train)] X_test_scaled = X_scaled[len(X_train):] ``` ✅ GOOD: Correct data splitting and transformation ```python from sklearn.preprocessing import StandardScaler from sklearn.model_selection import train_test_split import numpy as np X, y = np.random.rand(100, 5), np.random.rand(100) X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42) scaler = StandardScaler() # Fit only on training data X_train_scaled = scaler.fit_transform(X_train) # Transform both train and test using parameters learned from train X_test_scaled = scaler.transform(X_test) ``` ### 2.2. Reproducibility: Always Set `random_state` For any estimator or utility that involves randomness (e.g., `train_test_split`, `RandomForestClassifier`, `KMeans`), explicitly set the `random_state` parameter for reproducible results. ❌ BAD: Non-reproducible results ```python from sklearn.model_selection import train_test_split from sklearn.ensemble import RandomForestClassifier # Results will change on each run X_train, X_test, y_train, y_test = train_test_split(X, y) model = RandomForestClassifier() ``` ✅ GOOD: Reproducible results ```python from sklearn.model_selection import train_test_split from sklearn.ensemble import RandomForestClassifier # Use a fixed integer for random_state X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42) model = RandomForestClassifier(random_state=42) ``` ## 3. Performance Considerations - ### 3.1. Leverage Cython for Performance-Critical Components+ ### 3.1. Use Built-in Parallelism and Caching Before Custom Code - When extending scikit-learn with custom, performance-critical algorithms, use Cython and follow scikit-learn's internal conventions.+ Reach for scikit-learn's own performance features before writing custom numeric code. (Cython extensions using `sklearn.utils._typedefs` / `_openmp_helpers` are private APIs for contributing to scikit-learn itself, not for application code.) - - Disable bounds checking and wraparound for production code. - - Use `sklearn.utils._typedefs` for explicit type declarations. - - Use `sklearn.utils._openmp_helpers` for OpenMP routines. - - Prefer memoryviews over `cnp.ndarray` when possible.+ - Pass `n_jobs=-1` to estimators and `GridSearchCV`/`cross_validate` that support it. + - Prefer `HistGradientBoostingClassifier`/`Regressor` over `GradientBoosting*` on medium and large datasets. + - Cache expensive, repeated preprocessing with `Pipeline(..., memory="cache_dir")` during model selection. + - Keep sparse inputs sparse (`scipy.sparse`); do not call `.toarray()` on large one-hot or TF-IDF matrices. ```python- # Example of Cython directives and imports (in a .pyx file) - # distutils: language=c - # distutils: extra_compile_args=-fopenmp - # distutils: extra_link_args=-fopenmp - # cython: boundscheck=False - # cython: wraparound=False - # cython: cdivision=True+ from sklearn.ensemble import HistGradientBoostingClassifier + from sklearn.model_selection import GridSearchCV + from sklearn.pipeline import Pipeline + from sklearn.preprocessing import StandardScaler - from cython.parallel import prange - from sklearn.utils._typedefs cimport float64 - from sklearn.utils._openmp_helpers cimport _get_num_threads - - cdef void my_fast_function(float64[:] data, int n_threads) nogil: - cdef int i - with nogil: - for i in prange(data.shape[0], num_threads=n_threads, schedule='static'): - data[i] *= 2.0+ pipeline = Pipeline( + [("scaler", StandardScaler()), ("model", HistGradientBoostingClassifier(random_state=42))], + memory=".sklearn_cache", # transformers are fitted once per fold, not once per candidate + ) + search = GridSearchCV(pipeline, {"model__learning_rate": [0.05, 0.1]}, cv=5, n_jobs=-1) ``` ## 4. Common Pitfalls and Gotchas ### 4.1. Avoid Overfitting on Validation Data During Hyperparameter Tuning Use `GridSearchCV` or `RandomizedSearchCV` *with a pipeline* to ensure that cross-validation folds are correctly handled and preprocessing steps are refitted for each fold, preventing data leakage. ❌ BAD: Tuning hyperparameters outside a pipeline ```python from sklearn.preprocessing import StandardScaler from sklearn.linear_model import LogisticRegression from sklearn.model_selection import GridSearchCV, train_test_split import numpy as np X, y = np.random.rand(100, 5), np.random.randint(0, 2, 100) X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42) # Scaling outside of GridSearchCV means the scaler sees all training data # before cross-validation, leading to optimistic scores. scaler = StandardScaler().fit(X_train) X_train_scaled = scaler.transform(X_train) param_grid = {'C': [0.1, 1.0, 10.0]} grid_search = GridSearchCV(LogisticRegression(random_state=42), param_grid, cv=3) grid_search.fit(X_train_scaled, y_train) # Scores here are optimistically biased ``` ✅ GOOD: Tuning hyperparameters with a pipeline ```python from sklearn.pipeline import Pipeline from sklearn.preprocessing import StandardScaler from sklearn.linear_model import LogisticRegression from sklearn.model_selection import GridSearchCV, train_test_split import numpy as np X, y = np.random.rand(100, 5), np.random.randint(0, 2, 100) X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42) # Define the pipeline including preprocessing and the model pipeline = Pipeline([ ('scaler', StandardScaler()), ('logreg', LogisticRegression(random_state=42)) ]) param_grid = {'logreg__C': [0.1, 1.0, 10.0]} # Note the 'estimator__param' syntax grid_search = GridSearchCV(pipeline, param_grid, cv=3) grid_search.fit(X_train, y_train) # Fit the entire pipeline on X_train # Scores are realistic as preprocessing is done correctly within each fold ``` ## 5. Type Hints ### 5.1. Use Type Hints Extensively All function signatures, class attributes, and complex variable assignments must include type hints. This improves code readability, enables static analysis, and reduces bugs. ❌ BAD: Untyped function ```python def calculate_metric(y_true, y_pred): # What are y_true and y_pred? Lists, arrays, Series? return (y_true == y_pred).mean() ``` ✅ GOOD: Clearly typed function ```python import numpy as np- from typing import Union+ from numpy.typing import ArrayLike - def calculate_metric(y_true: Union[np.ndarray, list], y_pred: Union[np.ndarray, list]) -> float:+ def calculate_metric(y_true: ArrayLike, y_pred: ArrayLike) -> float: """Calculates accuracy for binary classification.""" y_true_arr = np.asarray(y_true) y_pred_arr = np.asarray(y_pred)- return (y_true_arr == y_pred_arr).mean()+ return float((y_true_arr == y_pred_arr).mean()) ``` ## 6. Virtual Environments ### 6.1. Isolate Project Dependencies with Virtual Environments - Always use a dedicated virtual environment (`venv`, `conda`, `poetry`, etc.) for each project. This prevents dependency conflicts and ensures consistent environments.+ Always use a dedicated virtual environment (`venv`, `uv`, `conda`, `poetry`, etc.) for each project. This prevents dependency conflicts and ensures consistent environments. ❌ BAD: Global `pip install` ```bash pip install scikit-learn pandas numpy # Pollutes global environment ``` ✅ GOOD: Project-specific virtual environment ```bash # Using venv python -m venv .venv source .venv/bin/activate pip install -r requirements.txt # Using conda- conda create -n my_ml_env python=3.9 scikit-learn pandas numpy+ conda create -n my_ml_env python=3.12 scikit-learn pandas numpy conda activate my_ml_env ``` ## 7. Packaging ### 7.1. Use `scikit-learn-contrib` Template for Shareable Estimators If you develop a custom estimator intended for public use or sharing across projects, leverage the `scikit-learn-contrib template`. It provides a robust structure, testing, and CI/CD setup, ensuring your estimator is fully compliant and easily consumable. ❌ BAD: Ad-hoc custom estimator distribution ```python # Just a .py file with a custom estimator, no proper packaging # Difficult for others to install, test, or integrate ``` ✅ GOOD: Structured project using `scikit-learn-contrib` ```bash # Follow the template to create a new project: # https://github.com/scikit-learn-contrib/project-template- # This provides setup.py, tests, documentation structure, etc.+ # This provides pyproject.toml, tests, documentation structure, etc. ``` ## 8. Testing Approaches ### 8.1. Implement Comprehensive Unit and Integration Tests For custom estimators, write unit tests to verify individual methods and use `sklearn.utils.estimator_checks.check_estimator` to ensure API compliance. For pipelines, write integration tests that cover the entire workflow. ❌ BAD: No tests or only manual verification ```python # No tests for MyCustomEstimator, relying on manual runs # to confirm it works. ``` ✅ GOOD: Automated testing with `check_estimator` ```python import pytest from sklearn.utils.estimator_checks import check_estimator+ from sklearn.utils.validation import check_is_fitted, validate_data from sklearn.base import BaseEstimator, TransformerMixin import numpy as np class MyCustomTransformer(BaseEstimator, TransformerMixin): def __init__(self, offset: float = 0.0): self.offset = offset def fit(self, X, y=None):- self._validate_data(X)+ validate_data(self, X) + self.is_fitted_ = True return self def transform(self, X):- self._validate_data(X, reset=False)+ check_is_fitted(self) + X = validate_data(self, X, reset=False) return X + self.offset # Test API compliance @pytest.mark.parametrize( "estimator", [MyCustomTransformer()] ) def test_all_estimators(estimator):- return check_estimator(estimator)+ check_estimator(estimator) # Add specific unit tests for logic def test_my_custom_transformer_offset(): X = np.array([[1, 2], [3, 4]]) transformer = MyCustomTransformer(offset=10) transformer.fit(X) transformed_X = transformer.transform(X) assert np.array_equal(transformed_X, np.array([[11, 12], [13, 14]]))- ```+ ```