Titanic Survival Prediction (Interpretable ML + Model Comparison)

A small, portfolio-ready machine learning project using the Titanic dataset.
The focus is interpretability + correct evaluation, not just accuracy.

What I tested (hypotheses)

Before modeling, I started with these assumptions:

• Women survived more than men
• Higher class → higher survival
• Younger passengers → higher survival
• Higher fare → higher survival (as a proxy for wealth)

Then I trained models to validate/refine those hypotheses with data.

---

Dataset

• File: Titanic-Dataset.csv
• Common Kaggle-style Titanic dataset used for ML practice.

---

Approach

1) Preprocessing

• Fill missing values:
  • Age → median
  • Embarked → mode
• Drop Cabin (too many missing values)
• Encode categorical variables:
  • Sex → binary (male=1, female=0)
  • Embarked → one-hot encoding with drop_first=True
    (baseline category becomes Embarked_C)

2) Model 1 — Logistic Regression (primary)

• Trained Logistic Regression for probability-based predictions: [ P(\text{Survived}=1 \mid X) ]
• Evaluated using:
  • Accuracy
  • Precision
  • Recall
  • F1
  • Confusion Matrix

3) Threshold tuning

• Default predict() uses threshold = 0.5
• I also tested threshold = 0.4 using predict_proba():
  • Improves recall and often F1 (tradeoff: more false positives)

4) Explainability

• Printed coefficients (log-odds)
• Converted coefficients → odds ratios
• Added:
  • % Change in Odds
  • Odds as % of Baseline

5) Model 2 — Random Forest (comparison)

• Trained RandomForestClassifier on the same features
• Compared metrics at:
  • default voting threshold
  • threshold = 0.4 (using probabilities)

---

Key insights (from Logistic Regression odds ratios)

• Sex and Pclass were the strongest predictors.
• Age and family-related features (SibSp/Parch) were secondary.
• Fare had minimal independent effect once class was included (often redundant with Pclass).

Summary conclusion:

“Survival was driven primarily by gender and class; age and family structure played secondary roles, while fare added little independent signal once class was accounted for.”

---

Project structure (suggested)

. ├── Titanic-Dataset.csv ├── titanic_model.py ├── README.md ├── requirements.txt └── LICENSE