AI and programming classes in Kirkcaldy

The learner downloads OpenStreetMap data over Kirkcaldy in sixteen tiles: 8,723 mapped buildings. Most, 5,917, carry no type and are left aside. Of the rest, 2,618 are tagged as homes of some kind and 170 as other buildings such as shops, garages, schools and churches. For each outline Python works out seven features: area, perimeter, number of corners, circularity, how fully it fills its tightest surrounding rectangle, how long and thin it is, and how many other buildings it touches, since terraces and semis share walls. The model is scored on held-back buildings over ten different random splits.

Because homes outnumber everything else by about fifteen to one, a model that always says "home" is right 93.9% of the time while being useless. So the fair score here is balanced accuracy: the average of the share of homes found and the share of other buildings found.

Balanced accuracy climbs to 85.6% at around 100 trees, beating a single decision tree at 82.7%, then slowly slips. Plain accuracy keeps rising all the way to 95.1%, because the extra trees chase the common case: the share of other buildings found falls from 81.2% to 72.5%. By then the model scores 99.3% on its own training buildings, a sign it is memorising rather than learning. With a learning rate of 1.0 the same thing happens faster and worse: balanced accuracy peaks at 82.6% after only 10 trees and ends at 80.2%, while plain accuracy reaches 96.3%. Small nudges and a sensible stopping point beat big steps and more trees.

Choosing the metric and the stopping point keeps Kirkcaldy teenagers in charge of the models they build, and learning to code is how they get that choice in 2026. The longer argument is in why teenagers should still learn to code in 2026.