Machine learning with scikit-learn

1. Concept and mental model

Machine learning uses examples to estimate a relationship between features and a target. In scikit-learn, a typical workflow separates data into training and test sets, fits a model on training data, and evaluates it on held-out data. Keep the test set out of feature selection, preprocessing decisions, and repeated model tuning; otherwise the reported test score becomes an optimistic estimate rather than an independent check.

2. How it works

Preprocessing must follow the same fitted transformation for training and production data. Use a pipeline to combine steps such as imputation, scaling, encoding, and model fitting. This reduces the chance of fitting a scaler to test data accidentally and makes the transformation reproducible. Select metrics based on the problem: accuracy can be misleading for imbalanced classes, while precision, recall, F1, or ranking metrics may better reflect the cost of errors.

3. Use it safely and effectively

Compare a model against a simple baseline and inspect errors by relevant subgroup or condition. A high score does not prove the model is fair, robust, or useful in deployment. Document data provenance, assumptions, and known limitations. Practice by training a small classifier with a pipeline and compare its evaluation to a baseline that predicts the majority class.

4. Syntax reference

from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=7)
model = make_pipeline(StandardScaler(), LogisticRegression())
model.fit(X_train, y_train)
print(model.score(X_test, y_test))

5. Guided walkthrough

Separate features from the target, split data before fitting transformations, and build preprocessing and model steps into a pipeline. Evaluate on held-out data that the model did not use during training, and compare against a simple baseline.

6. Common mistakes to avoid

• Fitting preprocessing on the full dataset before the train/test split, causing data leakage.
• Reporting accuracy alone when class imbalance or unequal error costs make it misleading.