scikit-learn-best-practices（scikit 学习最佳实践）

Scikit-learn Best Practices

Expert guidelines for scikit-learn development, focusing on machine learning workflows, model development, evaluation, and best practices.

Code Style and Structure

• Write concise, technical responses with accurate Python examples
• Prioritize reproducibility in machine learning workflows
• Use functional programming for data pipelines
• Use object-oriented programming for custom estimators
• Prefer vectorized operations over explicit loops
• Follow PEP 8 style guidelines

Machine Learning Workflow

Data Preparation

• Always split data before any preprocessing: train/validation/test
• Use train_test_split() with random_state for reproducibility
• Stratify splits for imbalanced classification: stratify=y
• Keep test set completely separate until final evaluation

Feature Engineering

• Scale features appropriately for distance-based algorithms
• Use StandardScaler for normally distributed features
• Use MinMaxScaler for bounded features
• Use RobustScaler for data with outliers
• Encode categorical variables: OneHotEncoder, OrdinalEncoder, LabelEncoder
• Handle missing values: SimpleImputer, KNNImputer

Pipelines

• Always use Pipeline to chain preprocessing and modeling
• Prevents data leakage by fitting transformers only on training data
• Makes code cleaner and more reproducible
• Enables easy deployment and serialization

from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.ensemble import RandomForestClassifier

pipeline = Pipeline([
    ('scaler', StandardScaler()),
    ('classifier', RandomForestClassifier(random_state=42))
])

Column Transformers

• Use ColumnTransformer for different preprocessing per feature type
• Combine numeric and categorical preprocessing in single pipeline

Model Selection and Tuning

Cross-Validation

• Use cross-validation for reliable performance estimates
• cross_val_score() for quick evaluation
• cross_validate() for multiple metrics
• Use appropriate CV strategy:

- KFold for regression - StratifiedKFold for classification - TimeSeriesSplit for temporal data - GroupKFold for grouped data

Hyperparameter Tuning

• Use GridSearchCV for exhaustive search
• Use RandomizedSearchCV for large parameter spaces
• Always tune on training/validation data, never test data
• Set n_jobs=-1 for parallel processing

Model Evaluation

Classification Metrics

• Use appropriate metrics for your problem:

- accuracy_score for balanced classes - precision_score, recall_score, f1_score for imbalanced - roc_auc_score for ranking ability

• Use classification_report() for comprehensive overview
• Examine confusion_matrix() for error analysis

Regression Metrics

• mean_squared_error (MSE) for general use
• mean_absolute_error (MAE) for interpretability
• r2_score for explained variance

Evaluation Best Practices

• Report confidence intervals, not just point estimates
• Use multiple metrics to understand model behavior
• Compare against meaningful baselines
• Evaluate on held-out test set only once, at the end

Handling Imbalanced Data

• Use stratified splitting and cross-validation
• Consider class weights: class_weight='balanced'
• Use appropriate metrics (F1, AUC-PR, not accuracy)
• Adjust decision threshold based on business needs

Feature Selection

• Use SelectKBest with statistical tests
• Use RFE (Recursive Feature Elimination)
• Use model-based selection: SelectFromModel
• Examine feature importances from tree-based models

Model Persistence

• Use joblib for saving and loading models
• Save entire pipelines, not just models
• Version control model artifacts
• Document model metadata

Performance Optimization

• Use n_jobs=-1 for parallel processing where available
• Consider warm_start=True for iterative training
• Use sparse matrices for high-dimensional sparse data
• Consider incremental learning with partial_fit() for large data

Key Conventions

• Import from submodules: from sklearn.ensemble import RandomForestClassifier
• Set random_state for reproducibility
• Use pipelines to prevent data leakage
• Document model choices and hyperparameters