Kaggle and Model-Free Prediction

Kaggle hosts the datasets and competitions we use for the rest of the course. This lecture covers how a competition is set up — what files you are given and what you submit back — and then establishes a model-free baseline: a prediction made from simple rules rather than from a trained model. Any model we build later has to beat that baseline to be worth its complexity.

Topics

• Rule-Based Prediction — two baselines on the Titanic data (predict everyone died; predict by gender), building the submission CSV with .loc[] and .apply(), and scoring both on Kaggle.
• The Heuristic Challenge — the in-class exercise: you invent the rules, on a booking-cancellation dataset nobody has told you anything about, and submit to a private leaderboard for this class only.

How a Kaggle Competition Works

A Kaggle competition normally provides at least three files:

1. the training data (for classification, the target is included)
2. the testing data (you need to predict the target)
3. the sample submission file (shows you the format your submission must follow)

Go to the Titanic competition page and download all three before class.

Datasets

• Titanic Competition — the competition we submit to throughout this lecture.
• titanic_dataset.zip — the same three Titanic files, bundled here if you cannot reach Kaggle.
• California Housing Dataset — the regression dataset we return to in later lectures.