Epicure: AI Embeddings Map Food Ingredient Geometry
A team of researchers has introduced Epicure, an innovative system built on three skip-gram ingredient embeddings. The foundation of this project is a comprehensive multilingual dataset comprising 4.14 million recipes, obtained from 11 distinct sources in seven languages, including English and Chinese. To enhance ingredient data, the study standardized raw ingredient strings into 1,790 entries. Additionally, the initiative created intricate graphs, featuring 203,508 ingredient connections and 80,019 ingredient-compound links with 2,247 compounds categorized into 15 groups. The research, which includes three variations of Metapath2Vec, was submitted to arXiv and is accompanied by 16 supplementary CSV files for in-depth exploration.
Key facts
- Epicure is a family of three skip-gram ingredient embeddings.
- Trained on 4.14 million recipes from 11 sources.
- Recipes span seven languages: English, Chinese, Russian, Vietnamese, Spanish, Turkish, Indonesian, German, and Indian-English.
- Raw ingredient strings normalized to 1,790 canonical entries via LLM-augmented pipeline.
- Constructed a 203,508-edge ingredient-ingredient NPMI graph.
- Constructed an 80,019-edge typed FlavorDB ingredient-compound graph with 2,247 typed compound nodes across 15 categories.
- Three Metapath2Vec variants: Cooc, Chem, and Core.
- Cooc walks co-occurrence graph only; Chem walks typed compound metapaths only; Core blends both.
- Submitted to arXiv; includes 16 ancillary CSV files.
- Models differ only in random-walk schema, sharing architecture and hyperparameters.
Entities
Institutions
- arXiv
- FlavorDB