Three overlapping styles in 215 snickerdoodle recipes

Each ratio is converted to a z-score (how many standard deviations a recipe sits from the mean), capped at 3 so one extreme recipe cannot claim a group, and k-means sorts the recipes into k groups, trying k from 2 to 4. Each k runs from 16 fixed starting points and keeps the tightest result, so the page is the same on every build. No group may hold fewer than 22 recipes.

The number of groups is the one with the best silhouette, which compares how close each recipe is to its own group with how close it is to the next one. By number of groups: 2 groups 0.18 · 3 groups 0.21 · 4 groups 0.18. At 0.25 or more the page calls the groups separate; from 0.18 to 0.25, overlapping; under 0.18, no clear styles.

A group's name lists the one to three ingredients where its recipes rank furthest from the middle of all 215, on average. “More” and “less” are relative to the whole dataset, and the names say nothing about texture or taste, which the ratios alone do not show. The median recipe of a group is, like the typical recipe on the overview, the middle of each ingredient separately, not one real recipe. The recipes listed under each group are the ones nearest its center, at most two from any one site.