AI and programming classes in Strabane

The learner downloads OpenStreetMap building outlines over Strabane and keeps the 2,939 on the UK side of the border, which the map also records. Of these, 1,155 are tagged as homes, 203 as other buildings such as shops, schools, sheds and garages, and 1,581 have no type. For every outline the program records its position, its floor area and a count of attached neighbours, because a terraced or semi-detached house shares mapped points with the house next door. The experiment hides most of the known labels, keeping only a small share, and asks two models to fill in the rest. A k-nearest-neighbours classifier learns from the kept labels alone. Label spreading builds a web linking every building, labelled or not, to its ten most similar neighbours, and lets the kept labels spread along it.

At every level, spreading labels through all the buildings beats learning from the labelled ones alone, by between 4.8 and 7.0 points. With only two non-home examples, the labelled-only model does no better than chance, 50.0%, while label spreading already edges ahead. The gain comes from the unlabelled buildings: they fill in the shape of the data, so a label in one terrace can flow along the whole street. It is not magic. Accuracy still climbs steeply as real labels are added, and if neighbouring buildings were unlike each other the spreading would mislead rather than help. We do not publish the model's guesses for the 1,581 untyped buildings, because nothing could check them.

A Strabane teenager who asks where a model's labels came from will not take its answers on trust, and writing the code is how that question becomes a habit. The longer argument is in why teenagers should still learn to code in 2026.