Applied Cases

A real dataset, a live leaderboard, and somewhere in the columns a trap someone left for you. Miss it and your model will look wonderful — brilliant validation score, confident submission — right up until the hidden holdout hands back a number you cannot explain. The trap has a boring name, leakage: a column that quietly knows the answer, a split that lets the future coach the past. Every experienced practitioner has been burned by it exactly once in public. This track is where you get burned in private instead, on purpose, with an instructor watching.

Nothing new is taught here — everything is used. Four cases, each a full argument from raw table to defensible conclusion. The first walks the honest end-to-end road: explore, clean, fit, validate, and write down a claim you would sign. The second drops the labels entirely and asks you to find real segments in real records — and to say how you know they are not artefacts of your own settings. The third is the leaderboard: a Kaggle-style challenge with an organizer's baseline to beat and a hidden holdout that scores what your pipeline actually generalises, not what your validation split flattered. The finale is a clinic — a parade of subtly broken models where you play diagnostician, separating leakage from overfitting from plain bad luck.

This is the capstone of the classical-ML pillar, and deliberately the track where the platform stops holding your hand: the methods come from ML Methods, the honest scoring from Evaluation, and the cases simply put you alone in the room with the data. What you take away is calibrated scepticism — the reflex, before believing any score including your own, to ask what the model could have cheated with and whether the evaluation would have caught it. Find the trap before you submit; the leaderboard remembers.