Iterative Development and Applications of Meituan Takeaway Food Knowledge Graph

1. Background

The food knowledge graph aims to describe entities and relationships of dishes in Meituan's takeaway service, enabling more accurate, rich, and personalized food recommendations. It aligns online takeaway dishes with offline dining data, providing a unified view for analysis.

2. Requirements and Challenges

Rapid growth of food items demands finer category tags, richer thematic attributes, and alignment between online and offline dish representations. Challenges include expanding the category taxonomy, extracting business‑specific thematic tags, and handling diverse dish expressions for entity alignment.

3. Knowledge Graph Iteration

3.1 Dish Category – Built a hierarchical taxonomy of over 300 categories using a BERT‑fine‑tuned classifier, replacing the earlier CNN+CRF model.

3.2 Classic Dish Tags – Leveraged standard dish names to aggregate category, flavor, ingredient, and cooking‑method tags, selecting top‑ranked dishes by sales and supply.

3.3 Health Meals – Defined low‑fat, low‑calorie meals and trained a hybrid Text‑CNN/Transformer model using keywords, dish names, merchant descriptions, and other textual features.

3.4 Dish Entity Alignment – Developed an algorithm that matches dish names across business lines by analyzing tokens, pinyin, prefixes/suffixes, and synonyms, producing a unified dish‑entity mapping.

4. Applications

4.1 Meal Combination – Dish Representation – Built an encoder‑decoder model with Multi‑Head Attention to predict compatible dishes for combo orders, improving conversion in various UI entry points.

4.2 Interactive Recommendation – Utilized standard dish tags to provide cross‑store, similar‑dish suggestions based on user preferences and real‑time behavior.

4.3 Search – Enhanced search accuracy and coverage by leveraging the high‑quality tags of the knowledge graph, leading to measurable improvements in user experience.

5. Future Plans

Explore scenario‑specific tag mining (e.g., festivals, weather, user groups) and multimodal models that combine text and images. Further research will focus on graph‑based recommendation techniques, such as random walks from user nodes to suggest relevant dishes.

6. References

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