Titanic Survival Prediction using Machine Learning

The Titanic dataset is the canonical first classification project, and most student builds waste it: they one-hot encode the raw columns, fit a model, and report an accuracy number nobody can learn from. The dataset rewards the opposite approach — careful feature engineering. Survival was driven by interactions the raw columns hide: class × sex ("women and children first" played out as 74.2% female vs 18.9% male survival), family size (small families beat both solo travellers and very large ones), and title (Master ≈ young boy, a strong survival signal). This project does the engineering properly: Title from names, FamilySize/IsAlone, deck from cabin, title-median age imputation, then a gradient-boosting classifier tuned with stratified cross-validation. The demo makes the model explainable — build a passenger, see the probability, and see exactly which factors moved the odds. Nothing is presented as a measured result before the training run; all metrics are labelled design targets.

1. The 891-row training set is loaded; the 418-row test set (labels withheld by Kaggle) is used only for the optional leaderboard submission path.
2. Features are engineered: Title parsed from Name, FamilySize = SibSp + Parch + 1, IsAlone flag, Deck letter from Cabin, LogFare, one-hot Embarked.
3. Missing values are handled explicitly: Age by median-per-title, Embarked by mode, Cabin by deck-or-missing indicator.
4. A gradient-boosting classifier is tuned via stratified 5-fold cross-validation over depth and learning rate.
5. The best configuration is evaluated once on a held-out 20% test split: accuracy, F1, ROC-AUC and confusion matrix go into the report.
6. The web demo embeds a simplified scoring function mirroring the engineered feature logic, with per-factor deltas so the prediction is explainable.

Tech stack:

• Python 3, scikit-learn (GradientBoosting, metrics, CV)
• NumPy, pandas, Matplotlib
• Jupyter Notebook (training & evaluation)
• HTML5 + JavaScript (survival predictor demo)
• Titanic dataset (Kaggle competition)

Dataset & model details

• Dataset: Titanic: Machine Learning from Disaster (Kaggle) — 891 training passengers, 12 raw columns (Pclass, Name, Sex, Age, SibSp, Parch, Ticket, Fare, Cabin, Embarked, plus PassengerId/Survived); 342 survived (38.4%). Test set: 418 rows, labels withheld.
• Task: Binary classification; input = 8 engineered features (Title, FamilySize, IsAlone, Deck, LogFare, Age, Pclass, Sex, Embarked), output = survival probability.
• Model: Gradient Boosting classifier, hyperparameters tuned with stratified 5-fold cross-validation.
• Metrics: Accuracy 82.1% (design target), F1 (survived) 0.74 (design target), ROC-AUC 0.87 (design target), confusion matrix on the 179-sample held-out test set. No metric is claimed as measured until the training run is executed for the order.

• [Real Kaggle Titanic dataset] 891 training passengers, 418 test rows, 38.4% survival — the "Machine Learning from Disaster" competition data.
• [Proper feature engineering] Title, FamilySize, IsAlone, Deck and LogFare engineered from the raw 12 columns, each justified in the report.
• [Principled missing-value handling] Age imputed by title-median, Embarked by mode, Cabin converted to deck/missing flag — no silent row-dropping.
• [Gradient boosting classifier] Handles mixed feature types and the non-linear class × sex interaction without manual interaction terms.
• [Stratified 5-fold cross-validation] Hyperparameter tuning that respects the small-n variance of 891 rows.
• [Explainable web demo] Build any passenger profile; get survival probability plus a per-factor contribution list showing what moved the odds.
• [Full evaluation] Accuracy, F1, ROC-AUC and confusion matrix on a held-out test set, reported once.

• Complete training & evaluation Jupyter notebook
• Trained gradient-boosting pipeline (.pkl) with preprocessing
• Explainable survival-predictor web demo
• Confusion matrix, ROC curve and feature-importance plots
• Project report PDF (feature-engineering rationale, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (boosting, cross-validation, leakage, imputation)

• 891 rows is small — cross-validation variance is real, and the report discusses it instead of hiding it.
• The dataset reflects 1912 demographics and a single disaster; it teaches ML method, not anything about modern survival.
• 82.1% accuracy is a design target, stated honestly — the report documents the actual achieved figures after training.
• The demo's live scorer is a simplified approximation for illustration; the notebook's trained model is the real deliverable.
• Kaggle test labels are withheld, so test-set evaluation uses a held-out split of the training data.