Beginner perfumer · 81
Every material has two blending lists, and they point in different directions
· 9 min read
The database's blending suggestions split into odour and flavour. 7,035 materials have both, and the two overlap only partly — the median intersection over union is 0.379, and over the shorter list it is 0.600. What is interesting is the direction: 24.3% of the odour half points at floral, while in the flavour half floral is only third at 10.9%, with fruity at 12.1% and citrus at 11.1% ahead of it. One database, one set of materials, two lists with different centres of gravity. So "what does this material go with" first requires knowing whether you are making a perfume or a flavour.
About 4 minutes.
This is the third article about the blending lists. The first two measured that the list is truncated at 60 per kind and that the odour half crosses families 73.6% of the time with a quarter pointing at floral.
This one deals with what both of those skipped: the list has two halves, and I had only looked at one.
Two lists
The kind column in the blenders table takes two values:
- odour: 336,574 rows
- flavour: 337,088 rows
Almost an even split. And 7,035 materials carry both.
How much they overlap
Treating each material's two lists as sets:
| Median | |
|---|---|
| Intersection ÷ union | 0.379 |
| Intersection ÷ shorter list | 0.600 |
The two lists share about a third of their union. Put another way, 60% of the shorter list also appears in the longer one.
Only one material out of 7,035 has no overlap at all, so the two lists are not talking past each other. They share a substantial base, and a substantial part differs.
Different centres of gravity
The real difference is direction. Setting the two destination distributions side by side:
| Rank | Odour half | Flavour half | ||
|---|---|---|---|---|
| 1 | Floral | 24.3% | Fruity | 12.1% |
| 2 | Balsamic | 12.0% | Citrus | 11.1% |
| 3 | Citrus | 9.7% | Floral | 10.9% |
| 4 | Fruity | 8.6% | Green | 7.3% |
| 5 | Herbal | 5.1% | Balsamic | 5.8% |
| 6 | Green | 4.0% | Herbal | 4.8% |
| 7 | Aldehydic | 3.2% | Spicy | 4.2% |
In the odour half, floral runs away with it — 24.3%, twice the runner-up.
In the flavour half, floral falls to third at 10.9%, with fruity and citrus ahead of it.
And the flavour half is far flatter. The odour half's leader takes 24.3%; the flavour half's takes 12.1% — no odour type holds even an eighth of the flavour list.
What this means
One: "what does this go with" needs a use first. The same material gets pushed toward floral for perfumery and toward fruity and citrus for flavour.
Two: the flavour half's suggestions are more dispersed. The odour list has an obvious centre; the flavour list does not. If you want combinations that are not the default, the flavour half's distribution is more even to begin with.
Three: 60% overlap makes cross-referencing worthwhile. A suggestion appearing in both is a stronger signal — it was listed under two different working contexts.
Four: the previous two articles need a qualifier. I wrote that the list points at floral. That holds only for the odour half. The flavour half does not behave that way, and I had not checked.
What to do
- Check which half you are reading. Odour for perfumery, flavour for food.
- Try the suggestions that appear in both first. A 60% overlap means the intersection is not random.
- To escape the floral default, read the flavour half. Its distribution is much flatter.
- The cap still applies. Each half is truncated at 60, so the combined 120 is two rankings, not one long list.
What this doesn't establish
- Overlap was computed on name strings. A material written differently in the two lists counts as non-overlapping, so 0.379 is a lower bound.
- The destination distributions use the
grpfield, and nothing guarantees the odour and flavour halves use the same classification vocabulary — the same methodological gap as last time, and worse here because I am comparing across the two halves. - "The flavour list is flatter" is an observation of the distribution. I computed no concentration index and ran no statistical test.
- Both halves carry many rows whose group field reads "no group found" (108,086 in the odour half, and some in the flavour half), all excluded from the distributions.
- I do not know how either list was generated, or whether they come from the same source.
- This article still evaluates no suggestion's quality. All three are about the shape of the list.