Measured pairing · from 63,696 recipes

What goes with milk fat?

Measured across 63,696 recipes, the 8 ingredients that co-occur with milk fat most disproportionately are lemon peel (×4.3 chance, n=49), brandy (×4.3 chance, n=27), orange peel (×3.4 chance, n=31), rum (×3.2 chance, n=30), cream (×2.4 chance, n=371), nutmeg (×2.3 chance, n=95), vanilla (×2.3 chance, n=313) and apricot (×2.3 chance, n=22). The score behind the ranking is positive pointwise mutual information — how many times more often a pair shares a recipe than random combination would predict — and n is the number of recipes containing both ingredients. Nothing is published below a floor of 20 recipes per pair and 100 per ingredient. Co-occurrence records what recipe authors actually do — it conflates deliciousness with popularity and tradition — so read it as ethnography, not judgment.

The measured partners — PPMI · ×chance · n

  1. lemon peel — PPMI 1.46 · ×4.3 chance · n=49 (lemon peel itself appears in 752 recipes)
  2. brandy — PPMI 1.45 · ×4.3 chance · n=27 (brandy itself appears in 420 recipes)
  3. orange peel — PPMI 1.22 · ×3.4 chance · n=31 (orange peel itself appears in 604 recipes)
  4. rum — PPMI 1.16 · ×3.2 chance · n=30 (rum itself appears in 621 recipes)
  5. cream — PPMI 0.88 · ×2.4 chance · n=371 (cream itself appears in 10,243 recipes)
  6. nutmeg — PPMI 0.85 · ×2.3 chance · n=95 (nutmeg itself appears in 2,686 recipes)
  7. vanilla — PPMI 0.83 · ×2.3 chance · n=313 (vanilla itself appears in 9,032 recipes)
  8. apricot — PPMI 0.83 · ×2.3 chance · n=22 (apricot itself appears in 638 recipes)

How this was measured

milk fat appears in 962 of the 63,696 recipes counted (5,296 parsed + 1,081 hand-curated + 57,319 from the dataset published with Ahn et al., 2011). PPMI = max(0, ln(count(A,B)·N / (count(A)·count(B)))), all counts over the same N; lift = e^PMI, which reads directly as "×chance." Staples (salt, sugar, flour, water) are excluded because sources under-record them, and pairs where one name contains the other are excluded as near-duplicates.

Read the numbers with their limits: the corpus is Western-skewed (90% of it is the largely North American and European recipe dataset published with Ahn et al.'s 2011 paper), so the absence of a pairing here proves nothing except who wrote the recipes down. PPMI favors the rare end even above the floor — a niche pair at n=22 can outrank a workhorse pair seen thousands of times, because common ingredients' huge counts divide the score down — so read PPMI and n together.

The full reference — 150 ingredients, top 8 partners each — lives at What actually goes with X.

Curious how these flavors fit your palate — take the short taste test →

Every number above is counted from recipes — pointwise mutual information over 63,696 of them — not chef interviews, model guesses, or reviews.