The Sale project"Also likes": association rules on Sale's shops
Group the town into baskets, measure support, confidence and lift, and then check the patterns against chance.
The learner downloads central Sale from the OpenStreetMap API in six tiles and keeps shops and a short list of everyday places such as cafés, restaurants, schools and the post office, leaving out categories this project does not need to discuss: 308 places in all. The first plan was to treat each street as a shopping basket, but only 14 streets had two or more kinds of place tagged, too few to learn from. So the town is cut into squares 200 m across instead, and each square's set of shop types becomes a basket: 68 squares have places, 30 of them two or more kinds.
Three numbers describe a rule such as "if a square has a clothes shop, it has a restaurant". Support is how many baskets contain both. Confidence is the share of clothes-shop squares that also have a restaurant. Lift compares that confidence with how common restaurants are anyway: a lift of 1 means no connection, above 1 means they turn up together more often than chance would suggest. The learner keeps pairs found in at least six squares, which gives 26 rules.
Strongest shop-pairings in Sale by 200 m square, our Python run on OpenStreetMap data, 29 September 2026| Rule | Squares with both | Confidence | Lift |
|---|
| Clothes shop, so restaurant | 7 | 1.00 | 1.76 |
| Hairdresser, so restaurant | 13 | 0.87 | 1.53 |
| Fast food, so convenience shop | 7 | 0.58 | 1.75 |
| Clothes shop, so hairdresser | 6 | 0.86 | 1.71 |
| Café, so convenience shop | 7 | 0.54 | 1.62 |
Confidence alone misleads. "Hairdresser, so restaurant" looks very strong at 0.87, but restaurants are the commonest kind of place, found in 17 of the 30 baskets, so a high confidence is almost guaranteed. Lift corrects for that popularity, and it ranks the rules differently: clothes shops and restaurants, fast food and convenience shops, clothes shops and hairdressers.
Then the honest check. The learner shuffles which kinds of place sit in which square 300 times, keeping every square's size and every type's count, and counts how many rules reach a lift of 1.5 or more by pure chance. The real town has 12 such rules. The shuffled towns have a median of 2 and a 95th percentile of 8, so Sale's patterns are stronger than chance usually produces. But one shuffle out of 300 also reached 12. With only 30 baskets, the evidence is real but modest, and no single rule should be treated as a law of shopping.
Ages 8 to 11
Record what is in five lunchboxes, find items that go together, then shuffle and see if it still happens.
Ages 11 to 15
Count pairs of shop types per square in Python and work out support and confidence.
Ages 15 and up
Compute lift for every rule, run the shuffle test and decide which rules deserve trust.
OpenStreetMap places, our rules
Shops and places are from OpenStreetMap and its contributors under the Open Database Licence. The baskets, rules and shuffle tests are our own work; no individual business is named or judged.
