Titanic Survival

The brief

Before you begin

Target

What you predict: Survived, 1 or 0.

Feature

A column you use to predict, such as Sex or Pclass.

Accuracy

The share of passengers you get right. Fine here because survivors are not rare (38%).

Baseline

The simplest sensible guess. A model must beat it to matter.

Feature engineering

Building a more useful column from raw ones, like pulling the title out of a name.

Overfitting

Learning the quirks of the training passengers instead of the pattern. Shows up as a good training score and a worse validation score.

Work through it, one step at a time

Each step is a short program you can run, read and edit. Press Run, look at what it prints, then open “What to notice”. The first run in a page loads Python and the data, which takes a few seconds.

1. Step 1 of 10

   Meet the passengers

   One row per passenger, with the answer in Survived. Check how common each answer is before anything else.

   Live Python

   Step 1: Meet the passengers

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") print(df.shape) print(df.head(3).T.to_string()) print() print("Survived (1) or not (0):") print(df["Survived"].value_counts()) print("Survival rate:", round(df["Survived"].mean() * 100, 1), "%")

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   891 passengers, 12 columns, and 38.4% survived. If you always guessed 'died' you would be right about 62% of the time, so a model has to do better than that.

2. Step 2 of 10

   Find what is missing, and what the gaps mean

   Age, Cabin and Embarked have holes. Before filling them, ask whether being missing tells you something.

   Live Python

   Step 2: Find what is missing, and what the gaps mean

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") print("Missing values per column:") print(df.isna().sum()[lambda s: s > 0]) print() print("Share missing: Age", round(df["Age"].isna().mean() * 100), "% | Cabin", round(df["Cabin"].isna().mean() * 100), "%") print() print("Survival rate when Cabin is known vs unknown:") print((df.groupby(df["Cabin"].notna())["Survived"].mean() * 100).round(1).to_string())

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   20% of ages and 77% of cabins are missing. Passengers with a known cabin survived 66.7% of the time against 30.0% without one. Cabin was recorded mostly for first class, so 'missing' is a clue about wealth, not random noise.

   Aha. A gap is information. Dropping or filling it without looking can throw the clue away.

3. Step 3 of 10

   Look at who survived

   Group averages show the pattern a model will learn. Look at one factor at a time, then two together.

   Live Python

   Step 3: Look at who survived

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") print("Survival rate by sex:") print((df.groupby("Sex")["Survived"].mean() * 100).round(1).to_string()) print() print("Survival rate by class:") print((df.groupby("Pclass")["Survived"].mean() * 100).round(1).to_string()) print() print("Survival rate by sex AND class:") print((df.groupby(["Sex", "Pclass"])["Survived"].mean() * 100).round(1).unstack().to_string())

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   Women survived 74.2% of the time, men 18.9%. First class survived 63.0%, third class 24.2%. Together they tell the story: first-class women survived 96.8% of the time, third-class men only 13.5%.

4. Step 4 of 10

   Split fairly, then set two baselines

   Split into train, validation and test with the same share of survivors in each. Then write the dumbest rules and score them.

   Live Python

   Step 4: Split fairly, then set two baselines

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") from sklearn.model_selection import train_test_split rest, test = train_test_split(df, test_size=0.2, stratify=df["Survived"], random_state=42) train, valid = train_test_split(rest, test_size=0.25, stratify=rest["Survived"], random_state=42) from sklearn.metrics import accuracy_score print("train / validation / test:", len(train), len(valid), len(test)) print() print("Baseline 1: nobody survives ->", round(accuracy_score(valid["Survived"], [0] * len(valid)), 3)) rule = (valid["Sex"] == "female").astype(int) print("Baseline 2: women survive, men do not ->", round(accuracy_score(valid["Survived"], rule), 3))

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   'Nobody survives' scores 0.618. 'Women survive, men do not' scores 0.798, one line of code and nearly 80%. Every model has to beat 0.798 to earn its complexity.

5. Step 5 of 10

   Build better columns

   Turn raw columns into ones a model can use: a 0/1 for sex, family size, the title hidden in the name, and a flag for 'age was missing'.

   Live Python

   Step 5: Build better columns

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") from sklearn.model_selection import train_test_split rest, test = train_test_split(df, test_size=0.2, stratify=df["Survived"], random_state=42) train, valid = train_test_split(rest, test_size=0.25, stratify=rest["Survived"], random_state=42) def prepare(frame, age_fill): X = pd.DataFrame(index=frame.index) X["Pclass"] = frame["Pclass"] X["female"] = (frame["Sex"] == "female").astype(int) X["Age"] = frame["Age"].fillna(age_fill) X["age_was_missing"] = frame["Age"].isna().astype(int) X["family_size"] = frame["SibSp"] + frame["Parch"] + 1 X["Fare"] = frame["Fare"] X["title"] = frame["Name"].str.extract(r", (\w+)\.", expand=False) X["title"] = X["title"].where(X["title"].isin(["Mr", "Mrs", "Miss", "Master"]), "Other") X["port"] = frame["Embarked"].fillna("S") return pd.get_dummies(X, columns=["title", "port"], dtype=float) age_fill = train["Age"].median() X_train, y_train = prepare(train, age_fill), train["Survived"] X_valid, y_valid = prepare(valid, age_fill).reindex(columns=X_train.columns, fill_value=0.0), valid["Survived"] X_test, y_test = prepare(test, age_fill).reindex(columns=X_train.columns, fill_value=0.0), test["Survived"] print("Median age used to fill gaps (from train only):", age_fill) print(X_train.shape) print(X_train.head(4).T)

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   14 columns, built from 7. The median age 29 is learned from train only. 'Master' (a young boy) and 'Mrs' will turn out to matter.

   Aha. The name column looks useless, yet its title carries age and status. Good features come from asking what a column means.

6. Step 6 of 10

   Train a model you can read

   Logistic regression gives each feature a weight. Positive pushes toward survival, negative away.

   Live Python

   Step 6: Train a model you can read

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") from sklearn.model_selection import train_test_split rest, test = train_test_split(df, test_size=0.2, stratify=df["Survived"], random_state=42) train, valid = train_test_split(rest, test_size=0.25, stratify=rest["Survived"], random_state=42) def prepare(frame, age_fill): X = pd.DataFrame(index=frame.index) X["Pclass"] = frame["Pclass"] X["female"] = (frame["Sex"] == "female").astype(int) X["Age"] = frame["Age"].fillna(age_fill) X["age_was_missing"] = frame["Age"].isna().astype(int) X["family_size"] = frame["SibSp"] + frame["Parch"] + 1 X["Fare"] = frame["Fare"] X["title"] = frame["Name"].str.extract(r", (\w+)\.", expand=False) X["title"] = X["title"].where(X["title"].isin(["Mr", "Mrs", "Miss", "Master"]), "Other") X["port"] = frame["Embarked"].fillna("S") return pd.get_dummies(X, columns=["title", "port"], dtype=float) age_fill = train["Age"].median() X_train, y_train = prepare(train, age_fill), train["Survived"] X_valid, y_valid = prepare(valid, age_fill).reindex(columns=X_train.columns, fill_value=0.0), valid["Survived"] X_test, y_test = prepare(test, age_fill).reindex(columns=X_train.columns, fill_value=0.0), test["Survived"] from sklearn.linear_model import LogisticRegression from sklearn.pipeline import make_pipeline from sklearn.preprocessing import StandardScaler from sklearn.metrics import accuracy_score model = make_pipeline(StandardScaler(), LogisticRegression(max_iter=1000)).fit(X_train, y_train) print("Logistic regression accuracy on validation:", round(accuracy_score(y_valid, model.predict(X_valid)), 3)) weights = pd.Series(model[-1].coef_[0], index=X_train.columns).sort_values() print() print("What it leans on (negative = less likely to survive):") print(weights.round(2).to_string())

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   Validation accuracy is 0.837. The largest weights are female (+0.96) and Pclass (-0.93), then family_size and Age. The model learned what the group averages showed, which is reassuring.

7. Step 7 of 10

   Try a random forest

   A random forest averages many decision trees. Compare it fairly with the logistic model on the same validation rows.

   Live Python

   Step 7: Try a random forest

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") from sklearn.model_selection import train_test_split rest, test = train_test_split(df, test_size=0.2, stratify=df["Survived"], random_state=42) train, valid = train_test_split(rest, test_size=0.25, stratify=rest["Survived"], random_state=42) def prepare(frame, age_fill): X = pd.DataFrame(index=frame.index) X["Pclass"] = frame["Pclass"] X["female"] = (frame["Sex"] == "female").astype(int) X["Age"] = frame["Age"].fillna(age_fill) X["age_was_missing"] = frame["Age"].isna().astype(int) X["family_size"] = frame["SibSp"] + frame["Parch"] + 1 X["Fare"] = frame["Fare"] X["title"] = frame["Name"].str.extract(r", (\w+)\.", expand=False) X["title"] = X["title"].where(X["title"].isin(["Mr", "Mrs", "Miss", "Master"]), "Other") X["port"] = frame["Embarked"].fillna("S") return pd.get_dummies(X, columns=["title", "port"], dtype=float) age_fill = train["Age"].median() X_train, y_train = prepare(train, age_fill), train["Survived"] X_valid, y_valid = prepare(valid, age_fill).reindex(columns=X_train.columns, fill_value=0.0), valid["Survived"] X_test, y_test = prepare(test, age_fill).reindex(columns=X_train.columns, fill_value=0.0), test["Survived"] from sklearn.ensemble import RandomForestClassifier from sklearn.linear_model import LogisticRegression from sklearn.pipeline import make_pipeline from sklearn.preprocessing import StandardScaler from sklearn.metrics import accuracy_score models = { "Logistic regression": make_pipeline(StandardScaler(), LogisticRegression(max_iter=1000)), "Random forest": RandomForestClassifier(n_estimators=200, max_depth=5, random_state=0), } for name, m in models.items(): m.fit(X_train, y_train) print(f"{name:20s} validation accuracy: {accuracy_score(y_valid, m.predict(X_valid)):.3f}")

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   The forest scores 0.854 against 0.837. A modest gain on only 178 passengers is within noise: with so few rows, one or two passengers change the score by a full percentage point.

8. Step 8 of 10

   Find where it is wrong

   Slice accuracy by group, and check which kind of mistake is more common.

   Live Python

   Step 8: Find where it is wrong

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") from sklearn.model_selection import train_test_split rest, test = train_test_split(df, test_size=0.2, stratify=df["Survived"], random_state=42) train, valid = train_test_split(rest, test_size=0.25, stratify=rest["Survived"], random_state=42) def prepare(frame, age_fill): X = pd.DataFrame(index=frame.index) X["Pclass"] = frame["Pclass"] X["female"] = (frame["Sex"] == "female").astype(int) X["Age"] = frame["Age"].fillna(age_fill) X["age_was_missing"] = frame["Age"].isna().astype(int) X["family_size"] = frame["SibSp"] + frame["Parch"] + 1 X["Fare"] = frame["Fare"] X["title"] = frame["Name"].str.extract(r", (\w+)\.", expand=False) X["title"] = X["title"].where(X["title"].isin(["Mr", "Mrs", "Miss", "Master"]), "Other") X["port"] = frame["Embarked"].fillna("S") return pd.get_dummies(X, columns=["title", "port"], dtype=float) age_fill = train["Age"].median() X_train, y_train = prepare(train, age_fill), train["Survived"] X_valid, y_valid = prepare(valid, age_fill).reindex(columns=X_train.columns, fill_value=0.0), valid["Survived"] X_test, y_test = prepare(test, age_fill).reindex(columns=X_train.columns, fill_value=0.0), test["Survived"] from sklearn.ensemble import RandomForestClassifier forest = RandomForestClassifier(n_estimators=200, max_depth=5, random_state=0).fit(X_train, y_train) out = valid[["Sex", "Pclass", "Survived"]].copy() out["predicted"] = forest.predict(X_valid) out["right"] = (out["predicted"] == out["Survived"]).astype(int) print("Accuracy by sex:") print((out.groupby("Sex")["right"].mean() * 100).round(1).to_string()) print() print("Accuracy by class:") print((out.groupby("Pclass")["right"].mean() * 100).round(1).to_string()) print() print("Share of mistakes that are 'predicted died, actually survived':", round(((out["predicted"] == 0) & (out["Survived"] == 1)).sum() / (out["right"] == 0).sum() * 100), "%")

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   Accuracy is similar for women (85.9%) and men (85.1%). Second class is easiest (94.4%) and third class hardest (82.8%). About 62% of the mistakes are survivors the model said would die.

   Aha. Overall accuracy hides which mistakes you make. Here, most are people who beat the odds.

9. Step 9 of 10

   Open the test set once

   Choose the forest, refit on train plus validation, and score the 179 test passengers one time. Compare with the one-line rule.

   Live Python

   Step 9: Open the test set once

   Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

   Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") from sklearn.model_selection import train_test_split rest, test = train_test_split(df, test_size=0.2, stratify=df["Survived"], random_state=42) train, valid = train_test_split(rest, test_size=0.25, stratify=rest["Survived"], random_state=42) def prepare(frame, age_fill): X = pd.DataFrame(index=frame.index) X["Pclass"] = frame["Pclass"] X["female"] = (frame["Sex"] == "female").astype(int) X["Age"] = frame["Age"].fillna(age_fill) X["age_was_missing"] = frame["Age"].isna().astype(int) X["family_size"] = frame["SibSp"] + frame["Parch"] + 1 X["Fare"] = frame["Fare"] X["title"] = frame["Name"].str.extract(r", (\w+)\.", expand=False) X["title"] = X["title"].where(X["title"].isin(["Mr", "Mrs", "Miss", "Master"]), "Other") X["port"] = frame["Embarked"].fillna("S") return pd.get_dummies(X, columns=["title", "port"], dtype=float) age_fill = train["Age"].median() X_train, y_train = prepare(train, age_fill), train["Survived"] X_valid, y_valid = prepare(valid, age_fill).reindex(columns=X_train.columns, fill_value=0.0), valid["Survived"] X_test, y_test = prepare(test, age_fill).reindex(columns=X_train.columns, fill_value=0.0), test["Survived"] import pandas as pd from sklearn.ensemble import RandomForestClassifier from sklearn.metrics import accuracy_score final = RandomForestClassifier(n_estimators=200, max_depth=5, random_state=0) final.fit(pd.concat([X_train, X_valid]), pd.concat([y_train, y_valid])) guess = final.predict(X_test) print("FINAL test result (look once)") print("Accuracy:", round(accuracy_score(y_test, guess), 3), "| passengers scored:", len(y_test)) print("Women-survive rule on the same passengers:", round(accuracy_score(y_test, (test["Sex"] == "female").astype(int)), 3))

   Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

   ▶ Run codeReset to starter
   What to notice

   Test accuracy is 0.821 against 0.777 for the women-survive rule. The model beats the rule by about four points, and the gap is smaller than what the validation numbers suggested. That shrinkage is normal and is exactly why the test set exists.

10. Step 10 of 10

    Write your submission file

    Kaggle wants PassengerId and Survived, nothing else. Build exactly that and download it.

    Live Python

    Step 10: Write your submission file

    Runs in your browser on your device. Nothing you type is sent anywhere, and nothing here changes your saved progress.

    Python code (editable)import pandas as pd df = pd.read_csv("data/titanic.csv") from sklearn.model_selection import train_test_split rest, test = train_test_split(df, test_size=0.2, stratify=df["Survived"], random_state=42) train, valid = train_test_split(rest, test_size=0.25, stratify=rest["Survived"], random_state=42) def prepare(frame, age_fill): X = pd.DataFrame(index=frame.index) X["Pclass"] = frame["Pclass"] X["female"] = (frame["Sex"] == "female").astype(int) X["Age"] = frame["Age"].fillna(age_fill) X["age_was_missing"] = frame["Age"].isna().astype(int) X["family_size"] = frame["SibSp"] + frame["Parch"] + 1 X["Fare"] = frame["Fare"] X["title"] = frame["Name"].str.extract(r", (\w+)\.", expand=False) X["title"] = X["title"].where(X["title"].isin(["Mr", "Mrs", "Miss", "Master"]), "Other") X["port"] = frame["Embarked"].fillna("S") return pd.get_dummies(X, columns=["title", "port"], dtype=float) age_fill = train["Age"].median() X_train, y_train = prepare(train, age_fill), train["Survived"] X_valid, y_valid = prepare(valid, age_fill).reindex(columns=X_train.columns, fill_value=0.0), valid["Survived"] X_test, y_test = prepare(test, age_fill).reindex(columns=X_train.columns, fill_value=0.0), test["Survived"] from sklearn.ensemble import RandomForestClassifier final = RandomForestClassifier(n_estimators=200, max_depth=5, random_state=0) final.fit(pd.concat([X_train, X_valid]), pd.concat([y_train, y_valid])) submission = pd.DataFrame({"PassengerId": test["PassengerId"].values, "Survived": final.predict(X_test)}) submission_csv = submission.to_csv(index=False) print("Your submission file has", len(submission), "rows. First rows:") print(submission.head(5).to_string(index=False))

    Press Ctrl+Enter (⌘+Enter on Mac) to run. Tab moves focus to the next control, as usual; use spaces to indent.

    ▶ Run codeReset to starter
    What to notice

    A 179-row file of 0s and 1s. On Kaggle itself you would train on all of train.csv and predict the competition's own test.csv.

Say the limits out loud

A result is only useful with its limits attached. Include these when you present it.

• This is a small, historical data set. A model that works here says little about modern risk.
• 'Women and children first' shows up in the data, so the model partly encodes who was given priority, not who was fittest.
• With only 891 passengers, small score differences are noise. Do not chase them.
• Prediction of survival is not an explanation. Do not infer that being in third class caused death.

Try next

• Try max_depth values from 2 to 8 in the forest and watch validation and training accuracy.
• Add a column for 'travelling alone' and see if it helps.
• Upload Kaggle's train.csv and compare your results with the numbers here.