Iterative Development and Applications of Meituan Takeaway Food Knowledge Graph

The article introduces the Meituan Takeaway Food Knowledge Graph, describing its overall architecture, tag taxonomy, and business-driven applications. It explains how the graph models four dimensions of food information: category tags, standardized dish names, basic attributes (ingredients, cuisine, flavor, cooking method, etc.), and theme attributes (e.g., signature dishes, merchant highlights).

1. Background

Knowledge graphs capture entities and relationships in the real world. In Meituan’s takeaway business, a food knowledge graph helps provide more accurate, rich, and personalized food services. Aligning online (takeaway) and offline (dine‑in) dish data enables cross‑scenario analysis.

2. Requirements and Challenges

Rapid growth of food items demands finer‑grained category tags, richer theme attributes, and alignment between online and offline dish representations. The graph must also handle dynamic merchant data (sales, supply, tags) and diverse dish naming conventions.

3. Knowledge‑Graph Iteration

Key iterative steps include:

Category Tags : Built from scratch with >300 hierarchical categories. Initially classified with a CNN+CRF model; later upgraded to a BERT‑pre‑trained fine‑tuned model for higher accuracy.

Standard Dish Names : Consolidate varied merchant inputs into canonical dish identifiers (e.g., "宫保鸡丁"). Over hundreds of thousands of standard dishes now cover most menu items.

Basic Attributes : Extract ingredient, cuisine, flavor, cooking method, and vegetarian/ non‑vegetarian tags per dish.

Theme Attributes : Capture business‑specific signals such as signature dishes, merchant specialties, and popularity metrics.

3.1 Health‑Meal Identification

Health meals are low‑fat, low‑calorie, high‑fiber dishes. A hybrid model combines Text‑CNN (for short titles) and Transformer (for longer descriptions) to classify health meals. Training data are enriched via keyword‑based sampling (e.g., "salad", "low‑calorie") and iterative augmentation.

3.2 Dish Entity Alignment

An alignment algorithm parses dish names into tokens (quantifiers, pinyin, prefixes/suffixes, substrings) and uses category recognition, standard‑name extraction, and synonym matching to unify entities across business lines. Example mappings: "碳烧鸽" ↔ "炭烧鸽", "番茄炒蛋" ↔ "西红柿炒蛋".

4. Applications

Combo Recommendation (Dish Representation) : An encoder‑decoder model with Multi‑Head Attention learns dish co‑ordering patterns. Encoder processes dish name, tags, and transaction statistics; decoder predicts the next dish in a combo, using beam search and a coverage mechanism.

Interactive Recommendation : Leverages standardized dish tags to provide cross‑store comparisons and personalized suggestions.

Search Enhancement : High‑coverage, accurate tag taxonomy improves keyword‑based food search, especially for newly added ingredients, cuisines, and functional tags.

5. Future Plans

Scene‑Specific Tags : Mining tags for festivals, weather, user groups (e.g., fitness, weight‑loss) and integrating multimodal signals (images, text).

Graph‑Based Recommendation Research : Combining user behavior graphs with the food knowledge graph for random‑walk‑based recommendations and deeper personalization.

6. References

[1] Kim Y. Convolutional neural networks for sentence classification. arXiv:1408.5882, 2014. [2] See A, Liu PJ, Manning CD. Get to the point: Summarization with pointer‑generator networks. arXiv:1704.04368, 2017. [3] Vaswani A et al. Attention is all you need. NeurIPS, 2017. [4] Hamilton W et al. Inductive representation learning on large graphs. NeurIPS, 2017.