The Hastings projectActive learning on 311 output areas
Pretend every label costs money, then compare choosing at random with letting the model choose.
Output areas are the smallest units the census publishes, a few hundred residents each. The learner downloads three tables from the Nomis API for all 311 in Hastings: type of home, cars per household and age. The label is whether at least half an area's households live in flats of any kind, which is true for 104 areas. The model sees three clues: the share of households without a car or van, the share of residents aged 65 or over, and the share aged 20 to 39. The first clue is strong. In the mostly-flats areas an average of 45.7% of households have no car, against 21.3% elsewhere.
Each experiment hides 93 areas as a test set and treats the other 218 as a pool of unlabelled areas. The model, a logistic regression, starts with just two labels, one of each kind. Then it gains labels one at a time. In the random version the next area is picked blindly. In the active version, called uncertainty sampling, the model is asked which unlabelled area it is least sure about, the one whose predicted probability is closest to a coin toss, and that area is labelled next. To avoid being fooled by one lucky split, the whole experiment runs 200 times with different random splits.
Mean test accuracy over 200 splits, 93 test areas each, our Python run, 28 September 2026| Labels used | Random choice | Uncertainty sampling |
|---|
| 5 | 75.7% | 74.6% |
| 10 | 79.9% | 79.7% |
| 20 | 82.1% | 83.1% |
| 40 | 83.9% | 84.8% |
| All 218 in the pool | 85.4% | 85.4% |
| Always answer "not mostly flats" | 66.6% | 66.6% |
The result is modest and honest. With only five labels, letting the model choose is slightly worse than picking at random, 74.6% against 75.7%; a model that has seen two examples has little idea what it does not know, a problem called the cold start. From about 20 labels the active version pulls ahead and stays there: at 40 labels it averages 84.8%, within 0.6 points of the 85.4% reached with all 218 labels. Comparing the two methods split by split, uncertainty sampling wins in 109 of the 200 splits at 40 labels, loses in 63 and ties in 28, so the gain is small but consistent rather than a lucky average.
One more check keeps the project grounded. A single rule, calling an area mostly flats when at least 39% of households have no car, scores 85.2% across all 311 areas. That rule was tuned on the very areas it is scored on, so it flatters itself a little, but it shows the task is easy and one clue does most of the work. Active learning is aimed mainly at tasks where labels are costly and the pattern is harder. The Hastings lesson is to measure the benefit rather than assume it.
Ages 8 to 11
Play twenty questions and discuss which question would split the choices most evenly.
Ages 11 to 15
Turn census counts into shares in Python and test the one-clue rule on real areas.
Ages 15 and up
Code uncertainty sampling, repeat over many splits and compare the methods pair by pair.
ONS counts, our experiment
The census counts for each output area are Office for National Statistics figures published through Nomis. The labels, the models, the 200 splits and every accuracy figure are our own work.
