cs.CYApr 2, 2026

Epicure: Multidimensional Flavor Structure in Food Ingredient Embeddings

Authors: Jakub Radzikowski, Josef Chen

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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

StageInputOutputKey technique
Raw FlavorGraph6,653—300D GNN embeddings
Step 1: Simplification6,653∼\sim 4,800LLM + Jaccard clustering
Step 2: Refinement∼\sim 4,800∼\sim 2,100LLM classification
Step 3: Ultra-refinement∼\sim 2,100∼\sim 1,400Domain-specific rules
Step 4: Manual QC∼\sim 1,400∼\sim 1,050Streamlit editor
Step 5: Consolidation∼\sim 1,0501,032ID reassignment
Table 1: Data curation pipeline summary. The pipeline achieves 6.4 ×\times consolidation while preserving embedding semantics and adding structured metadata.
CohortInclusion rulennUsed in
Curated (all)After curation pipeline1,032Secs. 3.1 – 3.4
Raw (all)All selected FlavorGraph nodes6,653LLM tagging (Sec. 2.2 )
Raw (categorized)Raw nodes with back-projected categories4,804Raw-vs-curated comparison (Sec. 3.7 )
Curated savouryCurated, flavor profile = savoury589Culture analysis (Sec. 3.6 )
Raw savouryRaw (categorized), flavor = savoury2,916Culture analysis (Sec. 3.6 )
USDA matchedCurated, matched to USDA712Chemical 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.
Dimensionncurn_{\text{cur}}Curated ρ\rhoρCV\rho_{\text{CV}}Raw ρ\rhoΔ\Delta
Scoville heat570.7570.20 †0.522++ 0.235
Climate latitude5090.5720.480.418++ 0.154
Umami1,0320.4990.480.390++ 0.110
Salty1,0320.4000.380.252++ 0.148
Sweet1,0320.3870.370.363++ 0.024
NOVA processing1,0320.3740.350.386−- 0.012
Table 3: Ordinal projection analysis: Spearman ρ\rho between LLM-assigned rank and embedding projection onto the low-to-high axis. ρCV\rho_{\text{CV}} : 10-fold cross-validated ρ\rho (axis defined on training folds only; 20 random repeats). All p<.001p<.001 except sour ( p=0.53p=0.53 , curated) and bitter ( p=0.42p=0.42 , curated), which fail the ordinal test; these two dimensions are instead characterized via extreme-level analysis (Table 4 ). nn : 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(SHU)\log_{10}(\text{SHU}) . Curated ρ=0.76\rho=0.76 , raw ρ=0.52\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, ρ\rho shown); sour and bitter use direct binary LLM labels (No/Yes, Cohen’s dd 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 ρ=0.37\rho=0.37 , raw ρ=0.39\rho=0.39 ). (b) Climate latitude: projection onto the tropical →\to subarctic axis, with per-zone violins ordered by latitude. Curated ρ=0.57\rho=0.57 , raw ρ=0.42\rho=0.42 . Top row: curated space. Bottom row: raw space. Axes are shared within each column.
 Direct binary ddOrdinal ρ\rho
TasteCur.dCVd_{\text{CV}}RawCur.Raw
Bitter0.910.580.640.030.04
Sour0.840.580.62−- 0.020.00
Table 4: Binary separation analysis for taste dimensions. “Direct binary dd ”: Cohen’s dd from independent yes/no LLM labels. dCVd_{\text{CV}} : 10-fold cross-validated dd (axis defined on training folds; 20 repeats). Ordinal ρ\rho from Table 3 shown for comparison. All p<.001p<.001 .
 SweetSaltySourBitterUmami
Sweet1.00−- 0.420.400.16−- 0.45
Salty 1.00−- 0.33−- 0.220.60
Sour  1.000.01−- 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 ±1\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 ( ρ\rho shown); sour and bitter use direct binary axes (Cohen’s dd shown). Solid lines indicate positive coupling ( cos⁡>0.35\cos>0.35 ); dashed lines indicate negative coupling ( cos⁡<−0.35\cos<-0.35 ).
 Binary dd 
DimensionCur.dCVd_{\text{CV}}RawΔd\Delta d
Hardness1.260.840.59++ 0.67
Moisture1.171.000.63++ 0.54
Fattiness1.020.830.48++ 0.54
Crunchiness0.930.660.48++ 0.45
Chewiness0.790.600.52++ 0.27
Table 6: Texture dimension analysis. Left: five binary dimensions, Cohen’s dd from direct yes/no LLM labels; dCVd_{\text{CV}} : 10-fold cross-validated (20 repeats). Right: viscosity (ordinal, restricted to liquids/semi-liquids). All p<.001p<.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, ρ\rho shown); the remaining five dimensions use direct binary LLM labels (No/Yes, Cohen’s dd shown). Numbers above violins show sample sizes.
Figure 6: Texture axis geometry via MDS. Five binary axes (No →\to Yes direction, dd shown) plus viscosity ordinal axis ( ρ\rho shown), positioned by inter-axis cosine dissimilarity. Solid/dashed lines indicate positive/negative coupling ( ∣cos⁡∣>0.3|\cos|>0.3 ).
DimensionChemical proxySourceLLM ρ\rhonnCur. ρ\rhoRaw ρ\rhoΔ\Delta
SweetTotal sugars (g/100 g)USDA0.394160.4090.402++ 0.01
SweetTotal sugars (mg/100 g)FooDB0.394300.4020.384++ 0.02
SaltySodium (mg/100 g)USDA0.406840.2860.164++ 0.12
UmamiL-Glutamic acid (mg/100 g)FooDB0.504940.2420.159++ 0.08
MoistureWater (g/100 g)USDA0.387120.3970.292++ 0.11
Table 7: Cross-validation of LLM-classified taste axes against measured chemical values. LLM ρ\rho : Spearman correlation from ordinal projection (Section 3.1 ). Cur./Raw ρ\rho : correlation of the embedding axis with laboratory-measured values in the curated and raw spaces respectively. All correlations are significant ( p<.001p<.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 ).
NutrientnnCur. ρ\rhoρCV\rho_{\text{CV}}ppRaw ρ\rhoΔ\Delta
Protein (g/100 g)7120.4730.406.8×10−416.8\times 10^{-41}0.304++ 0.17
Carbohydrates (g/100 g)7120.4710.421.4×10−401.4\times 10^{-40}0.339++ 0.13
Energy (kcal/100 g)7120.4690.402.7×10−402.7\times 10^{-40}0.308++ 0.16
Dietary fiber (g/100 g)6700.4430.371.6×10−331.6\times 10^{-33}0.272++ 0.17
Total fat (g/100 g)7120.4410.343.2×10−353.2\times 10^{-35}0.214++ 0.23
Table 8: Nutritional dimensions: Spearman ρ\rho between embedding projection (tercile-centroid axis) and USDA-measured nutritional values. Cur./Raw: curated (1,032) and raw (6,653) embedding spaces. ρCV\rho_{\text{CV}} : 10-fold cross-validated curated ρ\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<.001p<.001 ).
DimensionSubsetnnCur. ρ\rhoppRaw ρ\rhoTraining compounds
Total sugarsHub1560.4611.4×10−91.4\times 10^{-9}0.359None (sucrose, glucose, fructose absent)
Non-hub2740.3731.9×10−101.9\times 10^{-10}0.387
Glutamic acidHub1940.3051.5×10−51.5\times 10^{-5}0.216L-glutamic acid (1 compound)
Non-hub3000.1861.2×10−31.2\times 10^{-3}0.158
ProteinHub2790.4662.0×10−162.0\times 10^{-16}0.22812 amino acids (0.5% of edges)
Non-hub4330.4759.2×10−269.2\times 10^{-26}0.309
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. nn and pp are for the curated space.
Cuisinencurn_{\text{cur}}CuratedRawRaw (sub)Sub 95% CI
South Asian270.5020.3800.386[0.293, 0.515]
East Asian440.4320.3710.371[0.289, 0.485]
Southeast Asian240.4190.3270.332[0.266, 0.416]
Latin American440.4000.3130.313[0.264, 0.365]
Japanese460.3910.3310.328[0.278, 0.379]
Mediterranean1060.2880.2420.242[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=28W=28 , p=0.008p=0.008 ). Sub: raw subsampled to n=589n=589 (200 bootstrap iterations; 95% CI shown).
  PurityLift
CuisinennCur.RawRaw (sub)Cur.RawRaw (sub)
South Asian270.5890.3760.22712.8 ×\times21.9 ×\times13.4 ×\times
East Asian440.4660.2990.2376.2 ×\times11.9 ×\times9.7 ×\times
Japanese460.5390.4120.2776.9 ×\times14.0 ×\times9.4 ×\times
Latin American440.4750.3160.2366.4 ×\times8.2 ×\times6.1 ×\times
Mediterranean1060.4580.3890.3082.5 ×\times2.9 ×\times2.3 ×\times
Table 11: kk NN purity ( k=10k=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=589n=589 (200 bootstrap iterations). Curated purity exceeds the size-matched raw baseline for 7/7 cuisines (Wilcoxon p=0.008p=0.008 ).
Figure 8: 3D UMAP projection of 1,032 curated ingredients. Sweet (orange) and savoury (red) zones are delineated by 2.5σ2.5\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 1σ1\sigma covariance per cluster, visibly smaller in the curated space.
 Umami ∗∗∗Salty ∗∗∗Sour ∗∗∗Scoville ∗∗∗Latitude ∗∗∗NOVA ∗∗SweetBitter ∗∗∗
Japanese0.450.09−- 0.200.000.09−- 0.15−- 0.20−- 0.18
East Asian0.240.13−- 0.260.13−- 0.12−- 0.24−- 0.25−- 0.09
SE Asian0.140.12−- 0.160.20−- 0.27−- 0.24−- 0.18−- 0.07
South Asian−- 0.020.00−- 0.240.01−- 0.35−- 0.30−- 0.160.31
Latin American0.01−- 0.110.01−- 0.01−- 0.30−- 0.24−- 0.21−- 0.16
Mediterranean−- 0.05−- 0.07−- 0.040.000.05−- 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=10,000N=10{,}000 ; cuisine labels shuffled): ∗∗∗p<0.001{}^{***}p<0.001 , ∗∗p<0.01{}^{**}p<0.01 , ∗p<0.05{}^{*}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+0.5 (concentric ring at 0.5 = zero). Significance stars from permutation test ( N=10,000N=10{,}000 ). 13 of 14 axes discriminate significantly between cuisines ( p<0.01p<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 ∣ρ∣|\rho| ), sweet, salty, umami, scoville, viscosity, NOVA, latitude. Right: binary dimensions (Cohen’s dd ), 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 ∣ρ∣|\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 2.5σ2.5\sigma covariance ellipsoids showing the spatial extent of each region. The two poles (sweet, n=392n=392 ; savoury, n=581n=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=32 , min_dist =0.004=0.004 , metric == cosine .
ParameterValues testedRationale
n_neighbors15, 32 , 50, 100Controls local vs. global structure
min_dist0.001, 0.004 , 0.01, 0.05Controls cluster tightness
metriccosine (fixed)Matches embedding similarity metric
n_components3 (fixed)3D visualization
Appendix
Table 13: UMAP grid search configurations. The selected configuration (bold) maximizes mean intra-category cosine similarity in the 3D projection.
 Ordinal tasteSHUNOVATextureBinary tasteBinary texture  
IngredientSwSaUm  HardMoistSourBitterHardCrunchMoistFattyChewyCuisinesFP
sugarvhnn02harddrynonononononono–sw
soy saucenhh03liquidv.wetnonononoyesnonoJp,EA,SEAsv
miso pastelhh03gelwetnonononoyesnonoJpsv
cayennennn40k1softdrynonononononono–sv
wasabilnl01gelwetnonononoyesnonoJpsv
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).
More in this topic

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

Explore similar work

CardsList
May 21, 2026

Epicure: Navigating the Emergent Geometry of Food Ingredient Embeddings

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.

Jakub Radzikowski, Josef Chen

Graph Embedding
AskPDF
Aug 12, 2026

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

Cooking is a complex process that transforms raw ingredients into delicious and nutritious dishes, yet the recipes that encode this process remain largely free text; readable by people but not directly computable. Existing recipe collections capture fragments of this information, but no shared representation links a recipe's structured ingredient composition, its geo-cultural provenance, and its nutritional profile within a single queryable schema. We address this representation gap by formalizing a framework for recipe data structure that decomposes each recipe into typed ingredient entities, grounds those entities in a reference nutritional database, and annotates them with geo-cultural and dietary context. We present RecipeDB2, a structured compilation of 128,942 recipes with 35,474 ingredients from 32 regions and 99 countries. Ingredient phrases are parsed into seven culinary attributes using a transformer-based named-entity model; ingredients are linked to the USDA reference tables through a BERT embedding strategy (F1 = 87.90 on a manually adjudicated set of the 200 most frequent ingredients), yielding 148 nutritional parameters per mapped ingredient; a Random Forest classifier propagates 34 ingredient categories across the full vocabulary; and a deterministic, conservative rule set assigns each recipe a dietary style. Through RecipeDB2 (https://cosylab.iiitd.edu.in/recipedb2/), we demonstrate a scalable framework for making recipes computable, turning culinary heritage (long treated as an artistic rather than a quantitative object) into a data-driven analysis.

Mansi Goel, Sumit Bhagat, Saloni Srivastava +3

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

Knowledge RepresentationDocument Information Extraction
AskPDF
Apr 15, 2026

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

Humans routinely infer taste, smell, texture, and even sound from food images a phenomenon well studied in cognitive science. However, prior vision language research on food has focused primarily on recognition tasks such as meal identification, ingredient detection, and nutrition estimation. Image-based prediction of multisensory experience remains largely unexplored. We introduce FoodSense, a human-annotated dataset for cross-sensory inference containing 66,842 participant-image pairs across 2,987 unique food images. Each pair includes numeric ratings (1-5) and free-text descriptors for four sensory dimensions: taste, smell, texture, and sound. To enable models to both predict and explain sensory expectations, we expand short human annotations into image-grounded reasoning traces. A large language model generates visual justifications conditioned on the image, ratings, and descriptors. Using these annotations, we train FoodSense-VL, a vision language benchmark model to produce both multisensory ratings and grounded explanations directly from food images. This work connects cognitive science findings on cross-sensory perception with modern instruction tuning for multimodal models and shows that many popular evaluation metrics are insufficient for visually sensory inference.

Sabab Ishraq, Aarushi Aarushi, Juncai Jiang +1

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

VLM Evaluation
AskPDF