Advanced

Scikit-learn Best Practices

Deploy models to production, optimize performance for large datasets, debug common issues, and avoid the most frequent pitfalls in scikit-learn projects.

Production Deployment

Python

# Serve sklearn model with FastAPI
import joblib
from fastapi import FastAPI
from pydantic import BaseModel
import numpy as np

app = FastAPI()
model = joblib.load("model_pipeline.joblib")

class PredictionRequest(BaseModel):
    features: list[float]

@app.post("/predict")
def predict(req: PredictionRequest):
    X = np.array(req.features).reshape(1, -1)
    pred = model.predict(X)
    proba = model.predict_proba(X)
    return {"prediction": int(pred[0]), "probabilities": proba[0].tolist()}

Performance Optimization

Use n_jobs=-1
Most scikit-learn estimators support parallel execution. Set n_jobs=-1 to use all CPU cores for training and cross-validation.
Incremental Learning
For large datasets, use partial_fit() with SGDClassifier, MiniBatchKMeans, or MultinomialNB to train in chunks.
Sparse Matrices
Use scipy sparse matrices for high-dimensional, sparse data (e.g., text TF-IDF). Many algorithms support sparse input natively.
Feature Hashing
Use HashingVectorizer for text data to avoid storing a vocabulary in memory.

Common Pitfalls

Pitfall	Problem	Solution
Data leakage	Fitting scaler on full dataset	Always use pipelines with cross-validation
Target leakage	Feature contains future information	Review features for temporal dependencies
Ignoring class imbalance	Model predicts majority class	Use `class_weight="balanced"` or SMOTE
Wrong metric	Accuracy on imbalanced data	Use F1, ROC AUC, or PR AUC instead
Not setting random_state	Irreproducible results	Set random_state on models and splits
Overfitting to CV	Tuning too many params on CV	Use a held-out test set for final evaluation

Debugging Checklist

Check data shapes - Print X.shape and y.shape at each stage of your pipeline.
Inspect distributions - Plot feature distributions before and after preprocessing.
Learning curves - Use learning_curve() to diagnose underfitting vs overfitting.
Feature importance - Check model.feature_importances_ or use permutation importance.
Confusion matrix - Visualize with ConfusionMatrixDisplay to find systematic errors.

Quick Reference

Practice	Impact
Always use pipelines	Prevents data leakage, simplifies deployment
Set random_state everywhere	Reproducible experiments
Use cross_val_score, not train/test	More robust performance estimates
Start simple, add complexity	Baseline with LogisticRegression first
Version your data and models	Reproducibility in production
Monitor model performance	Detect data drift and degradation