ML Methods

You have labelled a hundred emails as urgent or not-urgent. Can the machine label the hundred-and-first? Usually yes — startlingly well, with methods that fit on a napkin. The better question is when it will embarrass you: on the sarcastic email, the one written in French, the one from a customer type it has never seen. Every classical method is a bet about what the world is like — that similar things sit near each other, that effects add up, that the future resembles the past — and it fails exactly where its bet fails. This track teaches the methods together with their bets.

The route runs from learning-with-labels to shipping-without-excuses. First the supervised workhorses, fitted and inspected one by one; then ensembles — why many mediocre models outvote one good one — and the unsupervised side, where there are no labels and the structure must be found rather than taught. Regression and GLMs put numbers rather than categories on the target, and a causal-inference module draws the platform's sharpest line: what prediction can never tell you about intervention, however good the fit. The back stretch is the unglamorous material working ML actually runs on — time series that leak the future into the past, missing values, broken assumptions — before model selection, interpretability, and a production-eye view of what happens after the notebook closes.

Each method is pushed until it breaks on purpose, because the breaking point is the lesson: a model you cannot embarrass on demand is a model you do not yet understand. That habit is also the best preparation for what follows — neural networks answer with scale and flexibility, but fitting, overfitting, validation and honest comparison mean exactly the same thing there, just with more zeros. Classical ML is still what most deployed systems run on; after this track you will know not only how to use each method, but how to argue for the one you chose.