P5 · Traditional Machine Learning

A model reports 99% accuracy on a fraud dataset. It is almost certainly worthless — if one transaction in a hundred is fraudulent, predicting "not fraud" every single time scores exactly that. The gap between a number that looks impressive and a model that is actually any good is the real subject of this pillar, and it is why evaluation gets a track of its own rather than a paragraph at the end.

The three tracks build in that order. ML Methods works the classical toolkit — regression through to clustering — but with an unusual instruction: each method is fitted, inspected, and then pushed until it breaks, because a model's assumptions are far easier to remember once you have watched them fail. Evaluation is the pillar's conscience: confusion matrices, ROC curves, calibration, and the specific habits of not fooling yourself, of which the accuracy trap above is only the most famous. Applied Cases then runs the whole thing end to end on real and imperfect tabular data, from a raw table to a conclusion you could defend in a meeting, finishing with a leaderboard-style case where the comparison is honest because everyone is scored the same way.