Project — Predict Titanic Survival

• Run the whole supervised pipeline on a real, messy dataset.
• Build a ColumnTransformer + model pipeline.
• Evaluate with cross-validation and a classification report.
• Interpret which factors mattered for survival.

import seaborn as sns
df = sns.load_dataset("titanic")    # built-in, no download
print(df[["survived","pclass","sex","age","fare","embarked"]].head())
print(df.isna().sum())              # age and others have gaps

   survived  pclass     sex   age     fare embarked
0         0       3    male  22.0   7.2500        S
1         1       1  female  38.0  71.2833        C
...
age          177   ← big gap to handle
embarked       2

1. LOAD     seaborn titanic
2. EXPLORE  survival rate by sex, class (a quick groupby)
3. SELECT   features: pclass, sex, age, sibsp, parch, fare, embarked
4. PREP     impute age/embarked, scale numerics, one-hot categoricals
5. TRAIN    pipeline + RandomForest, cross-validated
6. EVALUATE classification report + feature importances

Step 2 — explore first

print(df.groupby("sex")["survived"].mean().round(2))
print(df.groupby("pclass")["survived"].mean().round(2))

sex
female    0.74
male      0.19
pclass
1    0.63
2    0.47
3    0.24

Before training, you already expect sex and pclass to dominate. If your trained model disagrees, something is wrong.

Steps 3-4 — the pipeline pieces

num_cols = ["age", "sibsp", "parch", "fare"]
cat_cols = ["pclass", "sex", "embarked"]
# numeric: median impute + scale
# categorical: most_frequent impute + one-hot

# titanic.py — full ML pipeline
import seaborn as sns
import pandas as pd
from sklearn.model_selection import train_test_split, cross_val_score
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler, OneHotEncoder
from sklearn.impute import SimpleImputer
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report

df = sns.load_dataset("titanic")
y = df["survived"]
X = df[["pclass", "sex", "age", "sibsp", "parch", "fare", "embarked"]]

num_cols = ["age", "sibsp", "parch", "fare"]
cat_cols = ["pclass", "sex", "embarked"]

pre = ColumnTransformer([
    ("num", Pipeline([("imp", SimpleImputer(strategy="median")),
                      ("sc", StandardScaler())]), num_cols),
    ("cat", Pipeline([("imp", SimpleImputer(strategy="most_frequent")),
                      ("oh", OneHotEncoder(handle_unknown="ignore"))]), cat_cols),
])

model = Pipeline([("prep", pre),
                  ("clf", RandomForestClassifier(n_estimators=200, random_state=0))])

# cross-validated accuracy
cv = cross_val_score(model, X, y, cv=5)
print(f"CV accuracy: {cv.mean():.1%} ± {cv.std():.1%}")

# final fit + report on a held-out test set
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.2,
                                      stratify=y, random_state=0)
model.fit(Xtr, ytr)
print(classification_report(yte, model.predict(Xte),
                            target_names=["died", "survived"]))

# feature importances (names come out of the ColumnTransformer)
names = model.named_steps["prep"].get_feature_names_out()
imp = pd.Series(model.named_steps["clf"].feature_importances_, index=names)
print("\ntop factors:")
print(imp.sort_values(ascending=False).head(6).round(3))

Sample output

CV accuracy: 81.0% ± 2.4%

              precision    recall  f1-score   support
        died       0.83      0.86      0.84       110
    survived       0.77      0.72      0.75        69
    accuracy                           0.81       179

top factors:
cat__sex_male       0.27
num__fare           0.21
num__age            0.18
cat__pclass_3       0.10
num__sibsp          0.06

Read the diff

The model agrees with your exploration: sex and class/fare dominate. ~81% accuracy is solid for Titanic. The interpretable importances let you tell the story honestly — and notice that "survived" recall (0.72) is lower, meaning the model misses some survivors. That's where you'd focus next.

You ran a complete ML project: explore → prep (impute, scale, encode in a ColumnTransformer) → train (cross-validated forest) → evaluate (report + importances) → interpret. The model recovered the real history. This six-step shape is every supervised project you'll ever do. Next: the same shape, for predicting a number.