Online coding and Python classes in Wigan

The learner downloads the census occupation table for Wigan and for England. It gives nine groups, from managers, directors and senior officials to elementary occupations, each with a number in its name. The quickest way to feed this to a model, called label encoding, is to use those numbers directly. The learner does exactly that and averages them: 4.753 for Wigan, 4.387 for England. Then comes the test. The numbers are only labels, so the learner renumbers the same nine groups in alphabetical order and repeats the calculation. Wigan now scores 5.041 and England 5.047, so the gap has vanished and even reversed.

A result that changes when you rename the categories cannot be telling you anything about the categories. With label encoding, a model would also "learn" that elementary occupations are nine times managers, or that administrative work sits exactly halfway between two other groups, just because of the numbering. One-hot encoding avoids this. Each person becomes nine columns, all zero except a single 1 in their own group's column. Averaged over a place, those columns become the share of workers in each group, which is exactly the table above, and no group is bigger, smaller or between any other.

Read that way, the real picture is clear and does not depend on any numbering: Wigan has 4.1 points fewer professionals than England and 2.3 points more process, plant and machine operatives. The learner finishes by feeding both encodings to a tiny model that tries to tell Wigan workers from English ones, and watches the label-encoded version draw conclusions from the order of the list. The same care applies to every category an AI sees, from postcodes to product types.

Getting data preparation right is where good AI begins, and it is a strong reason for Wigan teenagers to keep learning to code in 2026. The longer argument is in why teenagers should still learn to code in 2026.