Epicure: Multidimensional Flavor Structure in Food Ingredient Embeddings

Jakub Radzikowski; Josef Chen

cs.CYApr 2, 2026

Epicure: Multidimensional Flavor Structure in Food Ingredient Embeddings

Authors: Jakub Radzikowski, Josef Chen

Abstract

A chef's intuition about flavor, texture, and cultural identity represents tacit knowledge that is difficult to articulate yet central to culinary practice. We show that this knowledge is already encoded in FlavorGraph's 300-dimensional ingredient embeddings, trained on recipe cooccurrence and food chemistry, and that it can be systematically recovered. An LLM-augmented curation pipeline consolidates 6,653 raw FlavorGraph ingredients into 1,032 canonical entries, substantially strengthening the recoverable structure. We identify at least fifteen independently classifiable dimensions spanning taste, texture, geography, food processing, and culture.

Figures & tables

Stage	Input	Output	Key technique
Raw FlavorGraph	6,653	—	300D GNN embeddings
Step 1: Simplification	6,653	$\sim$ 4,800	LLM + Jaccard clustering
Step 2: Refinement	$\sim$ 4,800	$\sim$ 2,100	LLM classification
Step 3: Ultra-refinement	$\sim$ 2,100	$\sim$ 1,400	Domain-specific rules
Step 4: Manual QC	$\sim$ 1,400	$\sim$ 1,050	Streamlit editor
Step 5: Consolidation	$\sim$ 1,050	1,032	ID reassignment

Table 1: Data curation pipeline summary. The pipeline achieves 6.4

\times

consolidation while preserving embedding semantics and adding structured metadata.

Cohort	Inclusion rule	$n$	Used in
Curated (all)	After curation pipeline	1,032	Secs. 3.1 – 3.4
Raw (all)	All selected FlavorGraph nodes	6,653	LLM tagging (Sec. 2.2 )
Raw (categorized)	Raw nodes with back-projected categories	4,804	Raw-vs-curated comparison (Sec. 3.7 )
Curated savoury	Curated, flavor profile = savoury	589	Culture analysis (Sec. 3.6 )
Raw savoury	Raw (categorized), flavor = savoury	2,916	Culture analysis (Sec. 3.6 )
USDA matched	Curated, matched to USDA	712	Chemical validation (Sec. 3.5 )

Table 2: Analysis cohorts. Each row defines a sample used in one or more analyses, with the inclusion rule and sections where it appears.

Dimension	$n_{\text{cur}}$	Curated $\rho$	$\rho_{\text{CV}}$	Raw $\rho$	$\Delta$
Scoville heat	57	0.757	0.20 †	0.522	$+$ 0.235
Climate latitude	509	0.572	0.48	0.418	$+$ 0.154
Umami	1,032	0.499	0.48	0.390	$+$ 0.110
Salty	1,032	0.400	0.38	0.252	$+$ 0.148
Sweet	1,032	0.387	0.37	0.363	$+$ 0.024
NOVA processing	1,032	0.374	0.35	0.386	$-$ 0.012

Table 3: Ordinal projection analysis: Spearman

\rho

between LLM-assigned rank and embedding projection onto the low-to-high axis.

\rho_{\text{CV}}

: 10-fold cross-validated

\rho

(axis defined on training folds only; 20 random repeats). All

p<.001

except sour (

p=0.53

, curated) and bitter (

p=0.42

, curated), which fail the ordinal test; these two dimensions are instead characterized via extreme-level analysis (Table 4 ).

n

: number of ingredients with valid tags.

Figure 1: Scoville heat scale: direct projection of spicy ingredients onto the heat axis (low-SHU to high-SHU tercile centroids) vs.

\log_{10}(\text{SHU})

. Curated

\rho=0.76

, raw

\rho=0.52

Figure 2: Taste dimension violin plots. Top row: curated space (1,032 ingredients). Bottom row: raw space (6,653 ingredients). Sweet, salty, and umami use ordinal projection (5-level none

\to

very_high axis,

\rho

shown); sour and bitter use direct binary LLM labels (No/Yes, Cohen’s

d

shown). Numbers above violins show sample sizes.

Figure 3: Ordinal projection violin plots. (a) NOVA processing level: the embedding axis separates unprocessed (NOVA 1) from ultra-processed (NOVA 4) ingredients (curated

\rho=0.37

, raw

\rho=0.39

). (b) Climate latitude: projection onto the tropical

\to

subarctic axis, with per-zone violins ordered by latitude. Curated

\rho=0.57

, raw

\rho=0.42

. Top row: curated space. Bottom row: raw space. Axes are shared within each column.

	Direct binary $d$			Ordinal $\rho$
Taste	Cur.	$d_{\text{CV}}$	Raw	Cur.	Raw
Bitter	0.91	0.58	0.64	0.03	0.04
Sour	0.84	0.58	0.62	$-$ 0.02	0.00

Table 4: Binary separation analysis for taste dimensions. “Direct binary

d

”: Cohen’s

d

from independent yes/no LLM labels.

d_{\text{CV}}

: 10-fold cross-validated

d

(axis defined on training folds; 20 repeats). Ordinal

\rho

from Table 3 shown for comparison. All

p<.001

	Sweet	Salty	Sour	Bitter	Umami
Sweet	1.00	$-$ 0.42	0.40	0.16	$-$ 0.45
Salty		1.00	$-$ 0.33	$-$ 0.22	0.60
Sour			1.00	0.01	$-$ 0.30
Bitter				1.00	$-$ 0.32
Umami					1.00

Table 5: Inter-axis cosine similarity between the five basic taste axes in the curated 300d space. Sweet, salty, and umami use ordinal (none

\to

very_high) axes; sour and bitter use binary (No

\to

Yes) axes. Values near 0 indicate orthogonal (independent) axes; values near

\pm 1

indicate aligned (confounded) axes.

Figure 4: Taste axis geometry via MDS projection of inter-axis cosine similarities. Sweet, salty, and umami use ordinal axes (

\rho

shown); sour and bitter use direct binary axes (Cohen’s

d

shown). Solid lines indicate positive coupling (

\cos>0.35

); dashed lines indicate negative coupling (

\cos<-0.35

Dimension	Cur.	$d_{\text{CV}}$	Raw	$\Delta d$
	Binary $d$
Hardness	1.26	0.84	0.59	$+$ 0.67
Moisture	1.17	1.00	0.63	$+$ 0.54
Fattiness	1.02	0.83	0.48	$+$ 0.54
Crunchiness	0.93	0.66	0.48	$+$ 0.45
Chewiness	0.79	0.60	0.52	$+$ 0.27

Table 6: Texture dimension analysis. Left: five binary dimensions, Cohen’s

d

from direct yes/no LLM labels;

d_{\text{CV}}

: 10-fold cross-validated (20 repeats). Right: viscosity (ordinal, restricted to liquids/semi-liquids). All

p<.001

Figure 5: Texture dimension violin plots. Top row: curated space (1,032 ingredients). Bottom row: raw space (6,653 ingredients). Viscosity uses ordinal projection (thin

\to

very_thick axis,

\rho

shown); the remaining five dimensions use direct binary LLM labels (No/Yes, Cohen’s

d

shown). Numbers above violins show sample sizes.

Figure 6: Texture axis geometry via MDS. Five binary axes (No

\to

Yes direction,

d

shown) plus viscosity ordinal axis (

\rho

shown), positioned by inter-axis cosine dissimilarity. Solid/dashed lines indicate positive/negative coupling (

|\cos|>0.3

Dimension	Chemical proxy	Source	LLM $\rho$	$n$	Cur. $\rho$	Raw $\rho$	$\Delta$
Sweet	Total sugars (g/100 g)	USDA	0.39	416	0.409	0.402	$+$ 0.01
Sweet	Total sugars (mg/100 g)	FooDB	0.39	430	0.402	0.384	$+$ 0.02
Salty	Sodium (mg/100 g)	USDA	0.40	684	0.286	0.164	$+$ 0.12
Umami	L-Glutamic acid (mg/100 g)	FooDB	0.50	494	0.242	0.159	$+$ 0.08
Moisture	Water (g/100 g)	USDA	0.38	712	0.397	0.292	$+$ 0.11

Table 7: Cross-validation of LLM-classified taste axes against measured chemical values. LLM

\rho

: Spearman correlation from ordinal projection (Section 3.1 ). Cur./Raw

\rho

: correlation of the embedding axis with laboratory-measured values in the curated and raw spaces respectively. All correlations are significant (

p<.001

). Of the compounds measured here, only L-glutamic acid is a node in FlavorGraph’s training graph; sucrose, glucose, fructose, sodium, and water are absent from the 1,561 training compounds (see Section 2.4 ).

Nutrient	$n$	Cur. $\rho$	$\rho_{\text{CV}}$	$p$	Raw $\rho$	$\Delta$
Protein (g/100 g)	712	0.473	0.40	$6.8\times 10^{-41}$	0.304	$+$ 0.17
Carbohydrates (g/100 g)	712	0.471	0.42	$1.4\times 10^{-40}$	0.339	$+$ 0.13
Energy (kcal/100 g)	712	0.469	0.40	$2.7\times 10^{-40}$	0.308	$+$ 0.16
Dietary fiber (g/100 g)	670	0.443	0.37	$1.6\times 10^{-33}$	0.272	$+$ 0.17
Total fat (g/100 g)	712	0.441	0.34	$3.2\times 10^{-35}$	0.214	$+$ 0.23

Table 8: Nutritional dimensions: Spearman

\rho

between embedding projection (tercile-centroid axis) and USDA-measured nutritional values. Cur./Raw: curated (1,032) and raw (6,653) embedding spaces.

\rho_{\text{CV}}

: 10-fold cross-validated curated

\rho

(axis defined on training folds, correlation measured on held-out fold; 20 random repeats). While aggregate macronutrient labels were not part of FlavorGraph’s training data, some constituent molecules (12 amino acids, 35 fatty acids) are among the 1,561 training compounds; the hub ablation (Table 9 ) quantifies the relative contribution of chemical vs. co-occurrence signal.

Figure 7: Nutritional dimensions: violin plots of embedding projections for ingredients grouped by USDA-measured tercile (Low/Mid/High). Clear monotonic separation confirms that the embedding encodes each nutrient. All correlations are highly significant (

p<.001

Dimension	Subset	$n$	Cur. $\rho$	$p$	Raw $\rho$	Training compounds
Total sugars	Hub	156	0.461	$1.4\times 10^{-9}$	0.359	None (sucrose, glucose, fructose absent)
Total sugars	Non-hub	274	0.373	$1.9\times 10^{-10}$	0.387	None (sucrose, glucose, fructose absent)
Glutamic acid	Hub	194	0.305	$1.5\times 10^{-5}$	0.216	L-glutamic acid (1 compound)
Glutamic acid	Non-hub	300	0.186	$1.2\times 10^{-3}$	0.158	L-glutamic acid (1 compound)
Protein	Hub	279	0.466	$2.0\times 10^{-16}$	0.228	12 amino acids (0.5% of edges)
Protein	Non-hub	433	0.475	$9.2\times 10^{-26}$	0.309	12 amino acids (0.5% of edges)

Table 9: Hub ablation: correlations for chemical-hub vs. non-hub ingredients in curated and raw spaces. Hub ingredients have direct compound edges in FlavorGraph’s training graph; non-hub ingredients received chemical information only via metapath propagation. Total sugars and glutamic acid use FooDB compound measurements projected onto LLM-defined taste axes; protein and total fat use USDA nutrient measurements with tercile-centroid axes. The “Training compounds” column indicates the chemical provenance: whether the measured molecule(s) are among the 1,561 FlavorDB training compounds.

n

and

p

are for the curated space.

Cuisine	$n_{\text{cur}}$	Curated	Raw	Raw (sub)	Sub 95% CI
South Asian	27	0.502	0.380	0.386	[0.293, 0.515]
East Asian	44	0.432	0.371	0.371	[0.289, 0.485]
Southeast Asian	24	0.419	0.327	0.332	[0.266, 0.416]
Latin American	44	0.400	0.313	0.313	[0.264, 0.365]
Japanese	46	0.391	0.331	0.328	[0.278, 0.379]
Mediterranean	106	0.288	0.242	0.242	[0.222, 0.264]

Table 10: Intra-cluster cosine similarity in the 300-dimensional embedding space (savoury ingredients only). Higher values indicate tighter cultural clusters. Curated exceeds raw for all 7 cuisines (Wilcoxon

W=28

p=0.008

). Sub: raw subsampled to

n=589

(200 bootstrap iterations; 95% CI shown).

		Purity			Lift
Cuisine	$n$	Cur.	Raw	Raw (sub)	Cur.	Raw	Raw (sub)
South Asian	27	0.589	0.376	0.227	12.8 $\times$	21.9 $\times$	13.4 $\times$
East Asian	44	0.466	0.299	0.237	6.2 $\times$	11.9 $\times$	9.7 $\times$
Japanese	46	0.539	0.412	0.277	6.9 $\times$	14.0 $\times$	9.4 $\times$
Latin American	44	0.475	0.316	0.236	6.4 $\times$	8.2 $\times$	6.1 $\times$
Mediterranean	106	0.458	0.389	0.308	2.5 $\times$	2.9 $\times$	2.3 $\times$

Table 11:

k

NN purity (

k=10

) in the 300-dimensional embedding space (savoury). Purity : fraction of same-cuisine neighbours. Lift : purity divided by random baseline. Raw (sub): subsampled to

n=589

(200 bootstrap iterations). Curated purity exceeds the size-matched raw baseline for 7/7 cuisines (Wilcoxon

p=0.008

Figure 8: 3D UMAP projection of 1,032 curated ingredients. Sweet (orange) and savoury (red) zones are delineated by

2.5\sigma

covariance ellipsoids around each pole centroid; transition-zone ingredients (black) lie between the poles. The black line marks the sweet

\leftrightarrow

savoury axis. An interactive 3D version is available at https://epicure-data.kaikaku.ai .

Figure 9: Projection of savoury ingredients onto the plane perpendicular to the sweet

\leftrightarrow

savoury axis (Figure 8 ). Left: curated space. Right: raw space. Labelled circles mark cuisine centroids; dashed ellipses show

1\sigma

covariance per cluster, visibly smaller in the curated space.

	Umami ∗∗∗	Salty ∗∗∗	Sour ∗∗∗	Scoville ∗∗∗	Latitude ∗∗∗	NOVA ∗∗	Sweet	Bitter ∗∗∗
Japanese	0.45	0.09	$-$ 0.20	0.00	0.09	$-$ 0.15	$-$ 0.20	$-$ 0.18
East Asian	0.24	0.13	$-$ 0.26	0.13	$-$ 0.12	$-$ 0.24	$-$ 0.25	$-$ 0.09
SE Asian	0.14	0.12	$-$ 0.16	0.20	$-$ 0.27	$-$ 0.24	$-$ 0.18	$-$ 0.07
South Asian	$-$ 0.02	0.00	$-$ 0.24	0.01	$-$ 0.35	$-$ 0.30	$-$ 0.16	0.31
Latin American	0.01	$-$ 0.11	0.01	$-$ 0.01	$-$ 0.30	$-$ 0.24	$-$ 0.21	$-$ 0.16
Mediterranean	$-$ 0.05	$-$ 0.07	$-$ 0.04	0.00	0.05	$-$ 0.22	$-$ 0.16	$-$ 0.09

Table 12: Cuisine centroid projections onto classified dimension axes (curated 300d space). Each value is the cosine dot product of the L2-normalized cuisine centroid with the unit axis vector. Significance assessed by permutation test (

N=10{,}000

; cuisine labels shuffled):

{}^{***}p<0.001

{}^{**}p<0.01

{}^{*}p<0.05

; no star indicates non-significant.

Figure 10: Cuisine profiles across classified dimensions. (a) Taste and heat dimensions (radar); (b) NOVA and latitude (bar chart); (c) Texture dimensions (radar). Radar panels show cosine projections shifted by

+0.5

(concentric ring at 0.5 = zero). Significance stars from permutation test (

N=10{,}000

). 13 of 14 axes discriminate significantly between cuisines (

p<0.01

or better); Sweet is non-significant.

Figure 11: Variant noise in the raw FlavorGraph data. (a) Mean pairwise cosine similarity among all variants of a consolidated ingredient, vs. the number of variants. Most groups cluster near the random baseline (0.28), indicating that variants of the “same” ingredient are barely more similar to each other than to random ingredients. (b) Minimum pairwise similarity: all 25 shown groups contain at least one pair with near-zero or negative cosine similarity.

Figure 12: Overview of curated vs. raw embedding space performance across all 14 validated dimensions. Left: ordinal dimensions (Spearman

|\rho|

), sweet, salty, umami, scoville, viscosity, NOVA, latitude. Right: binary dimensions (Cohen’s

d

), sour, bitter, hardness, crunchiness, chewiness, moisture, fattiness. The curated space outperforms the raw space on the majority of dimensions.

Figure 13: Curated vs. raw embedding performance on external laboratory measurements. Left: chemical cross-validation (Spearman

|\rho|

) against USDA nutrients and FooDB compounds, the two embeddings perform similarly on sweetness, but the curated space yields substantially higher correlations for salty–sodium, moisture–water, and umami–glutamic acid. Right: nutritional dimensions from USDA tercile centroids, the curated space consistently outperforms raw across all five macronutrients (protein, fat, fiber, calories, carbohydrates). Raw ingredients are matched independently to USDA (5,430/6,653, 81.6%) and FooDB (4,587/6,653, 68.9%) using the same three-layer pipeline.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 14: Toroid regions in the 3D UMAP projection, with

2.5\sigma

covariance ellipsoids showing the spatial extent of each region. The two poles (sweet,

n=392

; savoury,

n=581

) form dense clusters. The sweet pole is dominated by beverages, sweets, and fruits; the savoury pole by vegetables, meats, and spices.

Figure 15: UMAP projection from 300 dimensions to 3 dimensions. Left: curated space (1,032 ingredients). Right: raw space (6,653 ingredients, with back-projected categories). Each point is colored by the RGB average of its assigned category colors; marker shape varies by primary category. The curated space shows tighter, more distinct category clusters and a clearer toroidal structure. UMAP parameters: n_neighbors $=32$ , min_dist $=0.004$ , metric $=$ cosine .

Parameter	Values tested	Rationale
n_neighbors	15, 32 , 50, 100	Controls local vs. global structure
min_dist	0.001, 0.004 , 0.01, 0.05	Controls cluster tightness
metric	cosine (fixed)	Matches embedding similarity metric
n_components	3 (fixed)	3D visualization

Appendix

Table 13: UMAP grid search configurations. The selected configuration (bold) maximizes mean intra-category cosine similarity in the 3D projection.

Ingredient	Sw	Sa	Um			Hard	Moist	Sour	Bitter	Hard	Crunch	Moist	Fatty	Chewy	Cuisines	FP
	Ordinal taste			SHU	NOVA	Texture		Binary taste		Binary texture
sugar	vh	n	n	0	2	hard	dry	no	no	no	no	no	no	no	–	sw
soy sauce	n	h	h	0	3	liquid	v.wet	no	no	no	no	yes	no	no	Jp,EA,SEA	sv
miso paste	l	h	h	0	3	gel	wet	no	no	no	no	yes	no	no	Jp	sv
cayenne	n	n	n	40k	1	soft	dry	no	no	no	no	no	no	no	–	sv
wasabi	l	n	l	0	1	gel	wet	no	no	no	no	yes	no	no	Jp	sv

Appendix

Table 14: Spot-check of LLM classifications across 20 representative ingredients. Sweet/Salty/Umami: ordinal levels (n/l/m/h/vh = none/low/moderate/high/very_high). SHU: Scoville Heat Units. NOVA: processing level (1–4). Binary columns: direct yes/no LLM labels. Cuisines: distinctive cultural marker tags (Jp = Japanese, EA = East Asian, SEA = Southeast Asian, Med = Mediterranean, – = universal). FP: flavor profile (sw = sweet, sv = savoury, b = both, n = neutral).

Knowledge Representation

Knowledge representation studies how to encode facts, concepts, rules, context, and uncertainty so machines can interpret and reason over them.

See all 126 papers

Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings
May 21, 2026Jakub Radzikowski, Josef ChenGraph Embedding
A framework for recipe data structure with applications for culinary and nutritional insights
Aug 12, 2026Mansi Goel, Sumit Bhagat, Saloni Srivastava +3Knowledge Representation Document Information Extraction
Department of Computational Biology, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India · Infosys Centre for Artificial Intelligence, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India · Center of Excellence in Healthcare, Indraprastha Institute of Information Technology Delhi (IIIT-Delhi), New Delhi, 110020, India +2
FoodSense: A Multisensory Food Dataset and Benchmark for Predicting Taste, Smell, Texture, and Sound from Images
Apr 15, 2026Sabab Ishraq, Aarushi Aarushi, Juncai Jiang +1VLM Evaluation
College of Engineering and Computer Science, University of Central Florida, Orlando, FL, USA · College of Business Administration, University of Central Florida, Orlando, FL, USA · Institute of Artificial Intelligence, University of Central Florida, Orlando, FL, USA

May 21, 2026J/K move · Enter open · S save

Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings

Jakub Radzikowski, Josef Chen

Graph Embedding

Open paper AskPDF arXiv

We present Epicure, a family of three sibling skip-gram ingredient embeddings retrained from scratch on a multilingual recipe corpus. We aggregate 4.14M recipes from 11 sources spanning seven languages, English, Chinese, Russian, Vietnamese, Spanish, Turkish, Indonesian, German, and Indian-English, and normalise the raw ingredient strings to 1,790 canonical entries via an LLM-augmented pipeline. A 203,508-edge ingredient-ingredient NPMI graph and an 80,019-edge typed FlavorDB ingredient-compound graph, 2,247 typed compound nodes across 15 categories, seed three Metapath2Vec variants that share architecture and hyperparameters and differ only in the random-walk schema: Cooc walks the co-occurrence graph only, Chem walks the typed compound metapaths only, and Core blends both via injected ingredient-ingredient walks at controlled mixing, placing each model at a distinct point on the chemistry-vs-recipe-context spectrum.

Abstract

Knowledge Representation

Explore similar work

Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings

A framework for recipe data structure with applications for culinary and nutritional insights

FoodSense: A Multisensory Food Dataset and Benchmark for Predicting Taste, Smell, Texture, and Sound from Images