Home

No. 019 One shop, seven feeds

One shop, seven feeds

An imaginary shop of 1,200 handmade things, favorited by 2,500 shoppers who never existed. Seven recommenders sort it, side by side. Favorite a listing and every feed sorts itself again, each for its own reasons.

Most popular

Counts favorites. Every shopper sees the same ten.

  1. Striped cobalt bud vase, Rye Clay
  2. Ochre pair of taper candles, Fennel Candles
  3. Botanical sage pair of taper candles, Mallow Wax
  4. Speckled moss cylinder vase, Marrow Ceramics
  5. Botanical teal drop earrings, Blue Enamel
  6. Checked charcoal round rug, Paper Weavers
  7. Botanical blush tall print, Sorrel Press
  8. Striped cobalt bud vase, Lark Clay
  9. Plum paddle board, Lark Boards
  10. Speckled rust tall mug, Small Kiln

Trending and new

Counts this week’s favorites more than last month’s, and lifts what was just listed.

  1. Striped cobalt bud vase, Rye Clay
  2. Speckled ochre tall mug, Blue Kiln
  3. Checked cobalt area rug, Thistle Weavers
  4. Speckled sage paddle board, Linen Joinery
  5. Speckled moss cylinder vase, Marrow Ceramics
  6. Striped cobalt bud vase, Lark Clay
  7. Speckled rust espresso cup, Gull Stoneware
  8. Striped cream bottle vase, Slow Clay
  9. Cream jar candle, Blue Candles
  10. Speckled ochre tall mug, Sorrel Kiln

Similar items

Compares each listing with your favorites: craft, shape, surface, color, price, shop. It needs no other shoppers, and it narrows fast.

Shoppers like you

Looks at the shoppers who favorited what you did, and at what else they favorited. It finds the rug that goes with the mug, and it cannot see a listing nobody has favorited.

Learned taste

Gives every shopper and listing a place in a space learned from all the favorites, and looks near yours.

Explore

Starts from your taste, keeps repeats apart, and gives three places to listings it knows little about.

The blend

Takes candidates from the other feeds, scores them once, and sells two places.

Run the shop forward

Shoppers can only favorite what they are shown, so a feed also decides what gets favorited next. Here are seven copies of the shop, each run by one feed as if it were the only way in. Every day the same 100 shoppers come by in every copy, see that feed’s ten for them, and favorite what fits their taste. Every night the feed learns from the day.

As the run starts, the 120 most favorited listings hold 40% of all favorites. After 30 days:

  • Most popular earned 892 favorites and showed 20 of the 1,200 listings. 100% of those favorites went to listings already in the top tenth.
  • Trending and new earned 1,092 favorites and showed 22 of the 1,200 listings. 77% of those favorites went to listings already in the top tenth.
  • Similar items earned 1,601 favorites and showed 1,120 of the 1,200 listings. 18% of those favorites went to listings already in the top tenth.
  • Shoppers like you earned 2,962 favorites and showed 458 of the 1,200 listings. 86% of those favorites went to listings already in the top tenth.
  • Learned taste earned 3,335 favorites and showed 257 of the 1,200 listings. 93% of those favorites went to listings already in the top tenth.
  • Explore earned 2,501 favorites and showed 426 of the 1,200 listings. 92% of those favorites went to listings already in the top tenth.
  • The blend earned 3,057 favorites and showed 309 of the 1,200 listings. 84% of those favorites went to listings already in the top tenth.

How each feed decides

Most popular
Sorts by the number of favorites. It needs to know nothing about you, which is why it is what a new visitor gets. What is already seen gets seen more.
Trending and new
A favorite counts half as much for every week that has passed. A listing in its first week has nothing to count, so it is lifted: most on its first day, and more if its shop’s other listings are liked.
Similar items
Content-based filtering. Your favorites are averaged into a profile, attribute by attribute, and each listing is scored by the cosine between it and that profile. It can recommend a listing nobody has seen. It cannot surprise you.
Shoppers like you
Item-to-item collaborative filtering. Two listings are near when the same shoppers favorited both; each of your favorites votes for its neighbors. It never reads what a listing is, so it crosses from mugs to rugs when shoppers do.
Learned taste
Matrix factorization. From every shopper’s favorites, alternating least squares gives each listing eight numbers, so that listings liked by the same shoppers point the same way. Your own eight are fitted in the browser from your favorites, and a listing’s score is the product of the two.
Explore
Seven places go to your taste, each pick marked down for resembling your favorites and the picks above it. Three go to tries: listings with few favorites, from crafts you have not favorited, with a bonus that shrinks as favorites come in. A feed that never tries anything never learns anything new.
The blend
How a shop’s front page is put together. Candidates come from the other feeds; one score weighs taste, popularity and freshness; look-alikes are spaced out; then two places go to the shops that bid the most for a click, weighted by how well the listing fits you.

Nothing here is real. The listings, the shops and the shoppers are generated from a seed, and the favorites are simulated: each shopper has one or two of eight hidden tastes, and a taste spans crafts. The seven feeds are not simulated. They run in this page, on that data, as you click. When the shop runs forward, a shopper favorites a listing of their first taste about one time in ten in first place, less often lower down, and seldom anything outside their tastes.