Machine Learning Cheat Sheet

1 Which problem is it?

First two are supervised (you have labelled answers). Clustering is unsupervised. Know which before touching a model.

2 The workflow

• 1. Define the target & success metric.
• 2. Split train / test before anything else.
• 3. Explore & clean (on train only).
• 4. Engineer features.
• 5. Train a simple baseline.
• 6. Evaluate, tune, compare.
• 7. Ship & monitor.

Never let test data influence training — that's leakage, and it fakes great scores.

3 Split the data

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, random_state=42,
    stratify=y)   # keep class balance

Typical 80/20 or 70/30. random_state makes it reproducible. Keep the test set locked away until the very end.

4 Preprocess

from sklearn.preprocessing import StandardScaler, OneHotEncoder

# scale numbers, encode categories
StandardScaler().fit_transform(X_num)
OneHotEncoder(handle_unknown="ignore")

• Scale numeric features for distance/gradient models.
• Encode categoricals (one-hot or ordinal).
• Impute missing values (SimpleImputer).
• Fit on train, apply the same transform to test.

5 Pick a model

For tabular business data, start with a Random Forest or gradient boosting — rarely the wrong call.

6 Train & predict

from sklearn.ensemble import RandomForestClassifier

model = RandomForestClassifier(n_estimators=300,
                               random_state=42)
model.fit(X_train, y_train)

pred  = model.predict(X_test)
proba = model.predict_proba(X_test)[:, 1]

Every scikit-learn estimator follows the same fit → predict pattern — swap models with one line.

7 Score classification

99% accuracy is meaningless if 99% of cases are one class. Use the confusion matrix + F1/recall for fraud, churn, disease.

8 Score regression

from sklearn.metrics import mean_absolute_error, r2_score
mean_absolute_error(y_test, pred)
r2_score(y_test, pred)

9 Over/underfitting

• Overfit — great on train, poor on test. Memorised noise.
• Underfit — poor on both. Too simple.
• Fix overfit: more data, fewer features, regularise, simpler model.
• Trust cross-validation, not a single split:

from sklearn.model_selection import cross_val_score
cross_val_score(model, X, y, cv=5).mean()

10 Pipeline it

from sklearn.pipeline import make_pipeline

pipe = make_pipeline(StandardScaler(), model)
pipe.fit(X_train, y_train)

• A Pipeline bundles preprocessing + model — no leakage, one object to deploy.
• Tune with GridSearchCV / RandomizedSearchCV.
• Save with joblib.dump(pipe, "model.pkl").
• Monitor live: data drifts, and models go stale.