Don't Panic — Supervised Learning

Supervised learning shows a model examples with answers attached, then asks it to make the same kind of guess for a new case. It teaches a function by example rather than writing every rule by hand, which helps when a rule has too many corners for a spreadsheet to sulk through alone.

Each example has features, the input details, and a target, the known answer. A category such as spam is classification; a number such as a house price is regression. Give the model the wrong target, and it will work diligently on a question no one meant to ask.

Linear models add weighted inputs. Trees divide cases at feature thresholds. Ensembles combine models: random forests average independent trees, while gradient boosting corrects earlier errors. Nearest neighbors consult nearby labeled cases. Neural networks stack transformations when the relationship needs more flexibility. None has won the title of Universal Machine, chiefly because the universe declined to supply a test set.

The awkward part is evaluation. A training score says how well a model remembers examples, not how it handles new ones. Cross-validation repeatedly withholds part of the training data before a final untouched test set gives the honest estimate. For rare events, accuracy can look excellent while the model misses the cases anyone cared about, so precision, recall, and F1 matter.

Labels are the floor and ceiling of the arrangement. Biased historical labels reproduce their history. A changing world makes an old mapping stale even when code remains impeccably motionless. The Course tab gives the full map; Cheatsheet holds metrics and evaluation rules; Field Notes covers where tidy splits meet production data.