Build 12 indicative aroma heads up to confident - #270
Merged
Conversation
Adds structurally diverse, well-documented public-domain positives for 40 descriptors that were shipping as INDICATIVE — right less than half the time when they fire, because 11-30 positives cannot draw a boundary against 2,400 negatives. before 167 heads 94 confident-capable 73 indicative after 166 heads 106 confident-capable 60 indicative Two things were learned the hard way and are worth more than the numbers. PROBABILITY CALIBRATION DOES NOT WORK HERE, and it was tested first because it would have been free. CalibratedClassifierCV (isotonic and Platt) lifted `fruity` from 0.40 to 0.62 precision — it has 101 positives — while making `fishy` 0.46->0.33, `gardenia` 0.33->0.21 and `marine` 0.21->0.14 WORSE, and `cardamom`'s apparent 1.00 was a single true positive on ten examples. Isotonic regression needs data too, so it cannot rescue heads whose problem IS a lack of data; it only adds variance where variance is already the issue. Recorded on #262 so nobody spends a day on it. ADDING COMMON MOLECULES ACTIVELY HARMS A NARROW HEAD. The first pass gave `blackberry` ethyl butyrate, ethyl hexanoate, linalool and benzaldehyde — molecules present in essentially every fruit. Its AUROC fell 0.730 -> 0.626 and it dropped off the roster entirely. Stripping those and keeping only the distinctive markers (furaneol, methyl anthranilate, 2-heptanone) recovered it to 0.757. The same edit was applied across blueberry, guava, passionfruit, melon, mango and cassis. So "more positives" is too crude a rule, and this refines the method from #247 and #256: character-impact means DISTINCTIVE, not merely present. A molecule that occurs in every member of a category teaches the classifier nothing and dilutes what it already had. `fig` and `magnolia` had fallen under the 0.70 bar as the corpus shifted around them; rather than lose two heads they were given their documented character chemistry — sesquiterpenes and the creamy lactone for fig, the ocimene/anthranilate floral set for magnolia — and both came back (0.783 and 0.800), with fig now confident-capable at 0.60 precision. `blackberry` is the one casualty, at 0.696 against a 0.70 bar after two attempts. Lowering the bar to keep it would be exactly the self-flattery this project has been removing, so the bar stays and the head goes. It is tracked on #262 for the GNN work, where a better representation is the honest next lever. Corpus: aroma_supplement 2,159 associations across 165 descriptors (was 1,938 / 160); aroma_train 2,417 molecules. Downstream rebuilt: master_enrichment, profile_index (8,841 x 177) and flavor_map. Signed-off-by: Austin L. <86896075+rvnminers-A-and-N@users.noreply.github.com>
rvnminers-A-and-N
added
area:aroma
This was referenced Jul 30, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Progress on #262 (keeping it open — 60 heads still to go).
Adds structurally diverse public-domain positives for 40 descriptors, covering
citrus,almond,bready,earthy,cumin,oregano,thyme,dill,bergamot,violet,soapy,powdery, the berry/tropical set, the spice set, and the fatty/alcoholic ones.Two findings worth more than the numbers
Probability calibration does not work here — tested first, because it would have been free.
CalibratedClassifierCV(isotonic and Platt) liftedfruity0.40 → 0.62 precision (it has 101 positives) while makingfishy0.46 → 0.33,gardenia0.33 → 0.21 andmarine0.21 → 0.14 worse.cardamom's apparent 1.00 was one true positive on ten examples. Isotonic regression needs data too, so it can't rescue heads whose problem is missing data.Adding common molecules actively harms a narrow head. The first pass gave
blackberryethyl butyrate, ethyl hexanoate, linalool and benzaldehyde — present in essentially every fruit. AUROC fell 0.730 → 0.626 and the head dropped off the roster. Stripping those and keeping only distinctive markers (furaneol, methyl anthranilate, 2-heptanone) recovered it to 0.757. Same edit applied to blueberry, guava, passionfruit, melon, mango, cassis.That refines the #247/#256 method: character-impact means distinctive, not merely present. A molecule occurring in every member of a category teaches the classifier nothing and dilutes what it had.
Heads saved, and one lost honestly
figandmagnoliahad slipped under the 0.70 bar as the corpus shifted. Given their documented character chemistry — sesquiterpenes and the creamy lactone for fig, the ocimene/anthranilate set for magnolia — both returned (0.783 and 0.800), with fig now confident-capable at 0.60 precision.blackberryis the one casualty at 0.696 against a 0.70 bar, after two attempts. Lowering the bar to keep it would be exactly the self-flattery this project has spent five PRs removing, so the bar stays and the head goes. Tracked on #262 for the GNN work.Verification
master_enrichment,profile_index(8,841 × 177),flavor_map