scikit-learn（scikit 学习）

scikit-learn - Machine Learning in Python

A robust library for classical machine learning. It features a uniform API: all objects share the same interface for fitting, transforming, and predicting.

When to Use

• Classification: Detecting categories (Spam vs. Ham, Disease diagnosis).
• Regression: Predicting continuous values (House prices, Stock trends).
• Clustering: Grouping similar objects (Market segmentation, Image compression).
• Dimensionality Reduction: Reducing feature count while keeping info (PCA, Visualization).
• Model Selection: Comparing models and tuning hyperparameters (Cross-validation, Grid search).
• Preprocessing: Transforming raw data into features (Scaling, Encoding, Imputation).

Reference Documentation

Official docs: https://scikit-learn.org/stable/ User Guide: https://scikit-learn.org/stable/user_guide.html Search patterns: sklearn.pipeline.Pipeline, sklearn.model_selection, sklearn.ensemble, sklearn.preprocessing

Core Principles

The "Estimator" Interface

• Estimators: Implement fit(X, y). They learn from data.
• Transformers: Implement transform(X) (and fit_transform(X)). They modify data.
• Predictors: Implement predict(X). They provide estimates for new data.

Use scikit-learn For

• Tabular data (Excel-like, CSVs).
• Traditional ML (Random Forests, SVMs, Linear Models).
• Feature engineering and pipeline automation.
• Small to medium-sized datasets.

Do NOT Use For

• Deep Learning / Neural Networks (use PyTorch or TensorFlow).
• Natural Language Processing at scale (use spaCy or HuggingFace).
• Large-scale "Big Data" (use Spark MLlib or Dask-ML).
• Real-time streaming predictions (consider specialized inference engines).

Quick Reference

Installation

pip install scikit-learn

Standard Imports

import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split, cross_val_score, GridSearchCV
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer
from sklearn.metrics import classification_report, mean_squared_error

Basic Pattern - Train/Predict

from sklearn.ensemble import RandomForestClassifier

# 1. Prepare data
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

# 2. Instantiate and fit
model = RandomForestClassifier(n_estimators=100)
model.fit(X_train, y_train)

# 3. Predict and evaluate
predictions = model.predict(X_test)
print(classification_report(y_test, predictions))

Critical Rules

✅ DO

• Split before anything - Always use train_test_split before looking at data properties.
• Use Pipelines - Combine preprocessing and modeling to prevent data leakage.
• Scale your data - Models like SVM, KNN, and Linear Regression require feature scaling.
• Check for Imbalance - Use stratify=y in train_test_split for classification.
• Cross-Validate - Don't trust a single train/test split; use cross_val_score.
• Handle Missing Values - Use SimpleImputer or similar before fitting models.
• Standardize Categories - Use OneHotEncoder for nominal or OrdinalEncoder for ordinal data.

❌ DON'T

• Fit on test data - Never call .fit() or .fit_transform() on the test set.
• Use Categorical data as-is - Scikit-learn requires numerical input; encode strings first.
• Ignore Class Imbalance - Accuracy is misleading for imbalanced datasets; use F1-score or AUC.
• Overfit - Don't keep tuning hyperparameters until the test score is perfect.
• Ignore Random State - Set random_state for reproducibility during experiments.

Anti-Patterns (NEVER)

# ❌ BAD: Data Leakage (Fitting scaler on the whole dataset)
scaler = StandardScaler()
X_scaled = scaler.fit_transform(X) # Data from "future" test set leaks into training!
X_train, X_test, y_train, y_test = train_test_split(X_scaled, y)

# ✅ GOOD: Fit scaler only on training data
X_train, X_test, y_train, y_test = train_test_split(X, y)
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test) # Use training mean/std

# ❌ BAD: Repeating preprocessing manually
# (Error-prone and hard to maintain)

# ✅ GOOD: Use Pipelines (Automates everything safely)
pipe = Pipeline([
    ('scaler', StandardScaler()),
    ('model', RandomForestClassifier())
])
pipe.fit(X_train, y_train)

Preprocessing (sklearn.preprocessing)

Scaling and Encoding

from sklearn.preprocessing import StandardScaler, MinMaxScaler, OneHotEncoder

# Scaling numerical data
scaler = StandardScaler()
X_num_scaled = scaler.fit_transform(X_numeric)

# Encoding categorical data
encoder = OneHotEncoder(sparse_output=False, handle_unknown='ignore')
X_cat_encoded = encoder.fit_transform(X_categorical)

# Handling missing values
from sklearn.impute import SimpleImputer
imputer = SimpleImputer(strategy='median')
X_imputed = imputer.fit_transform(X_with_nan)

Column Transformer (The Pro Way)

from sklearn.compose import ColumnTransformer

numeric_features = ['age', 'salary']
categorical_features = ['city', 'job_type']

preprocessor = ColumnTransformer(
    transformers=[
        ('num', StandardScaler(), numeric_features),
        ('cat', OneHotEncoder(), categorical_features)
    ])

# Now use this in a pipeline
pipeline = Pipeline([
    ('prep', preprocessor),
    ('clf', LogisticRegression())
])

Classification

Common Algorithms

from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
from sklearn.ensemble import GradientBoostingClassifier

# Logistic Regression (Baseline)
log_reg = LogisticRegression(max_iter=1000)

# Support Vector Machine
svm = SVC(kernel='rbf', probability=True)

# Gradient Boosting
gb = GradientBoostingClassifier(n_estimators=100, learning_rate=0.1)

Regression

Common Algorithms

from sklearn.linear_model import Ridge, Lasso, ElasticNet
from sklearn.ensemble import RandomForestRegressor

# Regularized Linear Models
ridge = Ridge(alpha=1.0) # L2
lasso = Lasso(alpha=0.1) # L1

# Non-linear Regression
rf_reg = RandomForestRegressor(n_estimators=100, max_depth=10)

Model Evaluation

Metrics

from sklearn.metrics import accuracy_score, f1_score, roc_auc_score, r2_score, mean_absolute_error

# Classification
acc = accuracy_score(y_true, y_pred)
f1 = f1_score(y_true, y_pred, average='weighted')

# Regression
r2 = r2_score(y_true, y_pred)
mae = mean_absolute_error(y_true, y_pred)

Cross-Validation

from sklearn.model_selection import cross_val_score

scores = cross_val_score(pipeline, X, y, cv=5, scoring='f1_macro')
print(f"Mean F1: {scores.mean():.4f} (+/- {scores.std():.4f})")

Hyperparameter Tuning

Grid Search and Randomized Search

from sklearn.model_selection import GridSearchCV

param_grid = {
    'clf__n_estimators': [50, 100, 200],
    'clf__max_depth': [None, 10, 20],
    'clf__min_samples_split': [2, 5]
}

grid_search = GridSearchCV(pipeline, param_grid, cv=3, scoring='accuracy', n_jobs=-1)
grid_search.fit(X_train, y_train)

print(f"Best params: {grid_search.best_params_}")
best_model = grid_search.best_estimator_

Dimensionality Reduction

PCA (Principal Component Analysis)

from sklearn.decomposition import PCA

# Reduce to 2 components for visualization
pca = PCA(n_components=2)
X_pca = pca.fit_transform(X_scaled)

print(f"Explained variance ratio: {pca.explained_variance_ratio_}")

Clustering

K-Means and DBSCAN

from sklearn.cluster import KMeans, DBSCAN

# K-Means (Requires specifying K)
kmeans = KMeans(n_clusters=3, n_init='auto')
clusters = kmeans.fit_predict(X)

# DBSCAN (Density-based, finds K automatically)
dbscan = DBSCAN(eps=0.5, min_samples=5)
clusters = dbscan.fit_predict(X)

Practical Workflows

1. End-to-End Classification Pipeline

def build_and_train_model(X, y):
    # 1. Identify types
    num_cols = X.select_dtypes(include=['int64', 'float64']).columns
    cat_cols = X.select_dtypes(include=['object', 'category']).columns

    # 2. Setup Preprocessing
    preprocessor = ColumnTransformer([
        ('num', StandardScaler(), num_cols),
        ('cat', OneHotEncoder(handle_unknown='ignore'), cat_cols)
    ])

    # 3. Create Pipeline
    clf = Pipeline([
        ('preprocessor', preprocessor),
        ('classifier', RandomForestClassifier(random_state=42))
    ])

    # 4. Train
    X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y)
    clf.fit(X_train, y_train)

    return clf, X_test, y_test

# model, X_test, y_test = build_and_train_model(df.drop('target', axis=1), df['target'])

2. Custom Feature Engineering (Transformer)

from sklearn.base import BaseEstimator, TransformerMixin

class LogTransformer(BaseEstimator, TransformerMixin):
    def __init__(self, columns=None):
        self.columns = columns

    def fit(self, X, y=None):
        return self

    def transform(self, X):
        X_copy = X.copy()
        for col in self.columns:
            X_copy[col] = np.log1p(X_copy[col])
        return X_copy

Performance Optimization

Using n_jobs

# Use all CPU cores for training/tuning
model = RandomForestClassifier(n_jobs=-1)
grid = GridSearchCV(model, param_grid, n_jobs=-1)

Working with Large Data (partial_fit)

from sklearn.linear_model import SGDClassifier

# Online learning (incremental fit)
model = SGDClassifier()
for X_chunk, y_chunk in data_stream:
    model.partial_fit(X_chunk, y_chunk, classes=np.unique(y_all))

Common Pitfalls and Solutions

Imbalanced Classes

# ❌ Problem: Model predicts only the majority class
# ✅ Solution: Adjust class weights
model = RandomForestClassifier(class_weight='balanced')
# OR use SMOTE from imbalanced-learn library

Convergence Warnings

# ❌ Problem: "ConvergenceWarning: Liblinear failed to converge"
# ✅ Solution: Increase max_iter or scale data
model = LogisticRegression(max_iter=2000)
# Often solved by applying StandardScaler first!

Categorical Values in Test Set not in Train

# ❌ Problem: ValueError when unseen categories appear in test
# ✅ Solution: Use handle_unknown in OneHotEncoder
encoder = OneHotEncoder(handle_unknown='ignore')

Scikit-learn is the backbone of Python ML. Its API is so successful that many other libraries (XGBoost, LightGBM) mimic it.